Updated September 30, 2026

Best AI Models in 2026

Rankings, comparisons, and deep dives, updated monthly as new models ship.

Anthropic shipped Claude Sonnet 5.5 on September 28 at the same $2 / $10 per 1M tokens as Sonnet 5, and Artificial Analysis puts it at 56.0 on Intelligence Index v4.3.2, second only to Claude Opus 5.5 at 57.6 and ahead of Claude Fable 5.1 and GPT-6 Astra. It also leads Terminal-Bench 4.0 and LiveBench Coding, but Opus 5.5, which replaced Claude Opus 5 on September 22, keeps the coding crown: it now tops Arena's WebDev board, beats Sonnet 5.5 by 15 points on LiveBench Agentic Coding, and at max effort Sonnet 5.5 spends so many tokens that an index task costs $7.60 against Opus 5.5's $5.98. At its lower effort settings Sonnet 5.5 is the value pick, and it is the Claude model you get free on claude.ai.

On September 22, about ninety minutes after Opus 5.5, OpenAI answered with GPT-6 Sol and GPT-6 Luna, halving its API prices: Sol is $2 / $10 per 1M tokens and Luna $0.10 / $0.50, and neither is in ChatGPT's chat window yet. A week later, on September 29, it followed with GPT-6.1 Sol at the same $2 / $10: Artificial Analysis scores it 51.8, fifth among distinct models, and at $0.72 an index task it is the cheapest model in the top tier, though it too runs in ChatGPT Work and Codex rather than Chat. Arena has now rated Opus 5.5, and it went straight to first on the text, WebDev and creative-writing boards, while Claude Fable 5.1 keeps writing and accuracy on the boards those categories are decided on. This page ranks on measured scores, so a crown moves as soon as a model out-scores the holder, without waiting for the vote-based boards to catch up. Our guide to AI benchmarks explains how that index is built and why its version number matters.

SpaceXAI shipped Grok 4.7 on September 21 on a new and larger base model, and it places above Astra on both agentic boards Artificial Analysis publishes. The Intelligence Index reads 46.3 against Grok 4.6's 44.3 and the token price does not move, $2 / $6, but an index task costs $2.73 against $1.86, because 4.7 emits roughly twice the output. It launched on the API, Cursor and Grok Build; the Grok app still serves 4.6.

The open field has a new leader too: Xiaomi's MiMo-V2.6-Pro, published on September 21 under plain MIT, scores 46.3 and runs an index task for $0.13, the cheapest open model on this page. Below are the category winners for September 2026. Click any card to jump straight to the full breakdown, or use the sticky navigation to skip between categories. Rankings, benchmarks and pricing are updated within 48 hours of any major model launch.

September 2026 Category Winners

Writing

Claude Fable 5.1

#2 EQ-Bench Creative Writing and #3 LiveBench Language

In the top seven of all three prose boards, second on EQ-Bench and third on LiveBench Language. Best value: Claude Sonnet 5.5, free on claude.ai.

Deep dive →

Chat & Daily Assistant

GPT-5.6

ChatGPT default since July 9

The best assistant most people can actually open, balancing capability, speed, and reach.

Deep dive →

Images

ChatGPT Images 2.5

#1 and #2 on all four image boards we track

Its two models hold the top two places on every image board we track, Arena's and Artificial Analysis's alike, and it is the default on every ChatGPT tier including free. Best value: Flare, the faster of the two, at the same token price as GPT Image 2.

Deep dive →

Video

Gemini Omni 1.1 Flash

#1 Arena text-to-video (1516), 40-second scenes

Google's newest video model: scene extension to 40 seconds, 4K output, from $0.03 a second.

Deep dive →

Coding

Claude Opus 5.5

#1 Arena WebDev (1820) and #1 Intelligence Index (57.6)

Anthropic's September 22 flagship now leads Arena's WebDev board, beats Sonnet 5.5 by 15 points on LiveBench Agentic Coding and costs less per index task than Sonnet 5.5 at max effort. Best value: Claude Sonnet 5.5 at $2 / $10, which leads Terminal-Bench 4.0 (63.6%) and LiveBench Coding.

Deep dive →

Creativity

Grok 4.7

Fewest content restrictions + native real-time X

The pick for edgy, on-trend work: fewest guardrails and native X integration. Grok 4.7 is on the API, Cursor and Grok Build; the $30/month Grok app still serves 4.6.

Deep dive →

Accuracy & Research

Claude Fable 5.1

Highest factual accuracy Artificial Analysis has measured (67%)

Tops the boards that measure how often a model is simply wrong. Best value: Gemini 3.1 Pro at $0.52 a task with Google Search grounding.

Deep dive →

Problem Solving

GPT-6 Astra

#1 LiveBench Reasoning (92.65) and #1 ARC-AGI-2 (95%)

Took the hardest reasoning boards off GPT-5.6 Sol days after launching. Best value: GPT-6.1 Sol at $2 / $10, 0.02 behind it on LiveBench Reasoning.

Deep dive →

AI Agents

Gemini Spark

Always-on cloud agent from $19.99/month

Runs continuously in a Google Cloud VM, deeply integrated with Gmail, Docs, and Chrome.

Deep dive →

Want the top AI models without juggling separate subscriptions? Fello AI brings the leading models together in one native app for Mac, iPhone and iPad.

Download Fello AI

What's New in September 2026

Sep 29 OpenAI GPT-6.1 Sol reaches fifth on the Intelligence Index at $2 / $10, the cheapest per task in the top tier New

OpenAI released GPT-6.1 Sol at DevDay on September 29, 2026, a week after GPT-6 Sol, at the same $2 / $10 per 1M tokens with cached input halved to $0.10. OpenAI says it nearly matches GPT-6 Astra on agentic coding, computer use and professional work at a fifth of Astra's token prices, and the independent numbers put it close rather than level. Artificial Analysis scores it 51.8 on Intelligence Index v4.3.2 at max effort, fifth among distinct models behind Claude Opus 5.5 at 57.6, Claude Sonnet 5.5 at 56.0, Claude Fable 5.1 at 53.4 and Astra at 52.7, against 47.5 for GPT-6 Sol, and an index task costs $0.72, against $3.26 for Astra and $5.98 for Opus 5.5. LiveBench puts it second on Reasoning at 92.63, 0.02 behind Astra, third on Mathematics at 96.83 and second on Language at 90.13, but well outside the top ten on Coding at 80.7. It takes 56.1% on Terminal-Bench 4.0 against GPT-6 Sol's 43.9% and scores 1575 on GDPval-AA v2.1, above Astra's 1542. Arena, EQ-Bench and ARC Prize have not rated it yet. It is in ChatGPT Work and Codex for Plus, Pro, Business, Enterprise and Edu, not in Chat, and every GPT-6 tier bills 2x input and 1.5x output on the whole request once a prompt passes 272K tokens. OpenAI shipped no GPT-6.1 Astra beside it; the Wall Street Journal reported that its October release was cancelled after internal safety tests. The same event launched OpenAI dots, always-on agents that run on GPT-6 Astra, and a $500 ChatGPT Pro plan.

Sep 28 Anthropic Claude Sonnet 5.5 lands second on the Intelligence Index at 56.0, at Sonnet's $2 / $10 New

Anthropic released Claude Sonnet 5.5 on September 28, 2026, the second model in the Claude 5.5 family, at the same $2 / $10 per 1M tokens as Sonnet 5, with cache reads at $0.20. Anthropic says it runs more than 30% faster and costs up to 30% less per task, and the independent numbers back the capability claim: Artificial Analysis measures it at 56.0 on Intelligence Index v4.3.2 at max effort, second only to Claude Opus 5.5 at 57.6 and ahead of Claude Fable 5.1 at 53.4 and GPT-6 Astra at 52.7, against 38.2 for Sonnet 5. It takes Artificial Analysis's Terminal-Bench 4.0 run at 63.6% against Opus 5.5's 59.6%, LiveBench Coding at 91.4 against 89.3 and AutomationBench-AA at 71.3% against 69.5%, and it sits two Elo points behind Opus 5.5 on GDPval-AA v2.1, 1844 to 1846. The catch is effort. At max, Sonnet 5.5 spent 410 million tokens on the index and a task costs $7.60, against $5.98 for Opus 5.5 and $5.09 for Sonnet 5. At xhigh it scores 51.9 for $2.74 a task, close to GPT-6 Astra's 52.7 for less than Astra's $3.26, and at high 46.7 for $1.08. The Claude apps run it at Medium effort by default and the API at High. It is weaker where Opus 5.5 is strongest: 56.3 on LiveBench Agentic Coding against 71.7, and 32.3 on the hallucination-penalised AA-Omniscience index against 46.4. Arena has not ranked it yet. Anthropic says anyone can chat with Sonnet 5.5 on claude.ai, and higher-risk cybersecurity requests fall back to Sonnet 5, which is now a legacy model.

Sep 22 OpenAI GPT-6 Sol and GPT-6 Luna halve OpenAI's API prices, and land everywhere except ChatGPT's chat window New

OpenAI released GPT-6 Sol and GPT-6 Luna on September 22, 2026, about ninety minutes after Anthropic shipped Claude Opus 5.5, and the launch is an argument about cost per task rather than about the top of the board. Sol drops from $4 / $20 to $2 / $10 per 1M tokens and Luna from $0.20 / $1.20 to $0.10 / $0.50, with cached input reads discounted 90%, to $0.20 and $0.01. OpenAI labels both cuts 50%, but Luna's output falls 58.3%, and the comparison is against GPT-5.6 promotional pricing rather than list. Two spokespeople and OpenAI's Tibo Sottiaux all say the new rates are permanent. Where you can use them is the part every summary flattened: the models are in ChatGPT Work and Codex for Plus, Pro, Business, Enterprise and Edu, Free and Go users reach Luna in the desktop app, and OpenAI states plainly that they are "not yet available in Chat". Independently, Artificial Analysis scores Sol 47.5 on Intelligence Index v4.3.2 at max effort against GPT-5.6 Sol's 47.0, at $1.06 an index task against $1.99, with Terminal-Bench 4.0 at 43.9% against 39.9% and the hallucination rate down from 92% to 60%. Luna is level with GPT-5.6 Luna at 37.3 but runs an index task for $0.07 against $0.18, which makes it the cheapest model per index task on this page. ARC Prize has scored Luna, 59.3% on ARC-AGI-2 against the 5.6 build's 59.5%, and has published nothing for Sol. OpenAI's own evaluations are all cost arguments: on AutomationBench 1.0.6 Sol at xhigh effort takes 33.2% at $0.27 a task, and the closest rival row is Claude Fable 5.1 with an Opus 5 fallback at 31.4%, so the real lead is 1.8 points rather than the 6.3 the Claude Opus 5 row implies. OpenAI says Sol makes about half as many mistakes as GPT-5.6 Sol, "approaching Astra-level reliability at much lower cost", and repeats that Astra is still its best model across the board. Nothing here moves a crown: Sol and Luna are cheaper, not better. There is no GPT-6 Terra, and OpenAI has not said whether one is planned. Full detail in our GPT-6 Sol and Luna breakdown.

Sep 22 Anthropic Claude Opus 5.5 takes the Intelligence Index at 57.6 and undercuts Opus 5 on price New

Anthropic released Claude Opus 5.5 on September 22, 2026, the first model in a new Claude 5.5 family, and it is the largest single-day move this page has recorded. Artificial Analysis measures it at 57.6 on Intelligence Index v4.3.2, 4.3 points above Claude Fable 5.1 and the highest score it has published for any model, and it takes GDPval-AA v2.1 at 1846 against Fable 5.1's 1735, AA-Briefcase v1.1 at 1822 against 1678, Humanity's Last Exam at 61.4% against 59.1% and Artificial Analysis's own Terminal-Bench 4.0 run at 59.6% against GPT-6 Astra's 59.1% with it. Pricing goes down rather than up: $4 / $20 per 1M tokens against Opus 5's $5 / $25, cache reads 60% cheaper at $0.20 and cache writes at $5, with Anthropic's own testing putting a typical workload 40% cheaper than on Opus 5 and output more than 30% faster. A fast mode in Claude Code and the Claude Platform doubles the price for up to 2.5x the speed. On Anthropic's own harness the coding gap is wider still, 66.4% on Terminal-Bench 4.0 against the 57.9% OpenAI reports for Astra, plus 54.4% on FrontierCode v1.1 and 57.8% on CursorBench 4.0. Two things it does not win: Zapier's AutomationBench, where Astra scores 41.4% to its 40.0%, and Terminal-Bench-Science 0.1, where Astra scores 64.6% to its 58.7%. Read the margins with the caveat Anthropic prints itself, that the model "performs at the level of Claude Fable 5.1 on most work" and that "benchmark margins have become a less reliable guide to real-world differences". It is generally available on the Claude Platform as claude-opus-5-5 and on AWS, Google Cloud and Microsoft Azure, five-hour limits rise on Pro, Max and Team, and Claude Sonnet 5.5 and Claude Haiku 5.5 follow in the coming weeks. Because it matches Claude Mythos 5.1 in biology and cybersecurity it ships with Fable 5.1's safeguards and a new Life Sciences Verification Program for vetted labs. No Arena board has rated it yet. Full detail in our Claude Opus 5.5 breakdown.

Sep 21 SpaceXAI Grok 4.7 arrives on a larger base model at an unchanged $2 / $6, and costs 47% more per task New

SpaceXAI released Grok 4.7 on September 21, 2026 as its most capable model for coding and knowledge work. It uses a new and larger base model than Grok 4.6, with a longer reinforcement learning run weighted toward tasks that take many hours, and it was trained to understand the Grok Bot harness natively. The company publishes no parameter count. On its own figures it takes 46.3% on CursorBench 4.0 against Grok 4.6's 40.4% and GPT-5.6 Sol's 41.7%, 71.0% on DeepSWE v1.1 at high effort, 64.0% on EEBench, 38.0% on Terminal-Bench 4.0, 19.6% on the Harvey legal agent benchmark and 56.7% on HealthBench Professional. Independently, Artificial Analysis puts it at 46.3 on Intelligence Index v4.3.2 at high effort against Grok 4.6's 44.3, and at 1695 on GDPval-AA v2.1 and 1657 on AA-Briefcase v1.1, both above GPT-6 Astra's 1542 and 1569. The price does not move, $2 / $6 per 1M tokens with the whole request re-billed at $4 / $12 above 200,000 prompt tokens, and a fast variant runs at twice the output speed for twice the price. What does move is the cost per task: $2.73 against Grok 4.6's $1.86, because 4.7 emits roughly twice the output tokens getting to the same place, and xhigh effort buys no extra index points on either generation. On safety SpaceXAI reports the strongest refusal and jailbreak resistance it has tested, 62.4% on LatchBio's biosafety benchmark and 3.3% of risky dual-use prompts allowed through on its own HackerBench v0.3. Availability is developer-side: the Grok API, Cursor, Grok Build, third-party coding harnesses and cloud routers. The announcement says nothing about the consumer Grok app, which still serves Grok 4.6. Full detail in our Grok 4.7 breakdown.

Sep 21 Xiaomi Xiaomi's MiMo-V2.6-Pro is the highest-scoring open model ever measured, at $0.13 an index task under MIT New

Xiaomi published MiMo-V2.6-Pro on September 21, 2026 under plain MIT, with the weights on Hugging Face at XiaomiMiMo/MiMo-V2.6-Pro-RL on the day. It is a 1.02-trillion-parameter mixture of experts with 42 billion active and a 1M-token context. Artificial Analysis measures it, rather than estimating it, at 46.3 on Intelligence Index v4.3.2, which is the highest open score it has ever published: above GLM-5.3 at 44.8, above Kimi K3 at 43.6 and level with SpaceXAI's brand-new Grok 4.7. The cost is the other half of the story. It runs an Intelligence Index task for $0.13 at $0.435 / $0.87 per 1M tokens, which is cheaper per task than any other open model on this page, roughly half what GLM 5.3 Flash costs for four and a half points more. It also takes 34.8% on Terminal-Bench 4.0, 1673 on GDPval-AA v2.1, 49.4% on Humanity's Last Exam and 86.3% on AA-LCR, which puts it above Kimi K3 on three of those four. Two limits are worth stating plainly. No Arena board has rated it, so there is no human-preference evidence for it at all, and it was published four days ago, so it has no independent track record beyond Artificial Analysis. And 1.02 trillion parameters is a cluster, not a workstation, which is why GLM 5.3 Flash stays our recommendation for teams that actually want to host their own. It takes the open-weight crown on this page on measured score and cost.

Sep 15 TypeSafe TypeSafe leaves stealth with Jev, a model that returns typed decisions instead of text, at $0.042 per 1M tokens New

TypeSafe AI came out of two years of stealth on September 15, 2026 with $40 million in seed funding led by DCVC and a model that does not generate text at all. Jev is the first of what the company calls System One models: you send a block of state plus a fixed set of typed questions, and it answers all of them in one parallel pass, returning a choice, a score or a yes-no probability instead of a sentence. Because the answer space is defined before the call it cannot return a malformed result, which is where the launch post's line that Jev can't hallucinate comes from. It can still pick the wrong option, and TypeSafe's own documentation says calibration does not guarantee that an individual answer is correct. Pricing is $42 per billion input tokens, which is $0.042 per 1M, with output free, on a 64k context and text input only. The performance claims need care, because the company published four different ones in three days: from the press release's under 100 milliseconds and up to 100x faster to the homepage's 193.6x faster and 444.6x cheaper, each measured against a different rival, with its own footnote calling the big numbers the higher end of real world gains. Its workflow eval scores against the average of GPT-6 Astra and Fable 5.1 rather than ground truth, and it publishes no public benchmark scores at all by choice. Two independent early-access tests on September 18 measured roughly 6x cheaper and 8x faster against a cheap model with reasoning off, and the calibration held up. Nothing here moves a pick on this page: Jev has no consumer surface and is waitlisted early access. Full detail in our System One models explainer.

Sep 10 DeepSeek DeepSeek retires V4-Flash, replaces it with V4.1-Flash and cuts the Flash rate by about a third New

DeepSeek retired DeepSeek-V4-Flash on September 10, 2026 and put DeepSeek-V4.1-Flash in its place. The model name is now deepseek-flash; the legacy deepseek-v4-flash and deepseek-v4-flash-vision-exp names still resolve, but the models behind them have been retired and the requests are served by V4.1-Flash and billed at the Flash price. That price is a cut of roughly a third: $0.15 / $0.60 per 1M tokens off-peak and $0.30 / $1.20 at peak, against $0.22 / $0.66 and $0.44 / $1.32 for the build it replaces, with cache hits at $0.003 off-peak. The peak windows are unchanged, 01:00 to 04:00 and 06:00 to 10:00 UTC on weekdays, with every other hour at half. V4-Pro is untouched at $0.66 / $1.98 off-peak, and DeepSeek has said it will keep serving it past its September 14 retirement date. The new model is a 552-billion-parameter mixture-of-experts on a causal encoder-decoder design that activates 8 billion parameters at prefill and 16 billion at decode, takes a 1M-token context, reads images natively, and went up on Hugging Face under MIT the same day. Artificial Analysis scores it 39.5 on Intelligence Index v4.3 against 34.5 for the V4-Flash 0731 build it replaces, at $0.27 an index task. It does not retake the price-performance pick: GLM 5.3 Flash scores 41.9 at $0.25 a task. Full detail in our DeepSeek pricing guide.

Sep 3 OpenAI GPT-6 Astra ships at $10 / $50, and within a day it is on Plus and Pro and leading the coding boards New

OpenAI launched GPT-6 Astra on September 3, 2026, settling the argument that ran all year: this is GPT-6, not another point release inside the GPT-5 line. Company president Greg Brockman told a press briefing "Welcome to the AGI era." The independent numbers are more measured than the framing. Artificial Analysis scores Astra 61 on Intelligence Index v4.1.1 at max effort, exactly level with GPT-5.6 Sol, five points behind Claude Fable 5.1 at 66 and two behind Claude Opus 5 at 63, and it charges $10 / $50 per 1M tokens against Sol's $4 / $20, which works out 75% more expensive per index task than the model it replaces. The stronger result is on the Coding Agent Index, where Astra in Codex scores 67, level with Claude Opus 5 and Claude Fable 5 in Claude Code and behind Claude Fable 5.1's 70, and it gets there on one third of Sol's tokens and one fifth of Opus 5's, which puts it on the cost-efficiency frontier. It also halves hallucination on AA-Omniscience, from 92% to 51% at max effort, while gaining four points of accuracy, and adds roughly 80 Elo on AA-Briefcase and six points on Humanity's Last Exam. Against that it drops about 80 Elo on GDPval-AA v2 and regresses two to three points on customer support, SciCode and long-context reasoning. Two caveats travel with the launch numbers: the 98.6% on ARC-AGI-3 is an action-efficiency score against a human baseline rather than a solve rate, measured on a harness OpenAI itself has shown can triple the result, and Epoch AI discloses that OpenAI funded FrontierMath and holds exclusive access to part of it. Availability is the real constraint. Astra is the first model OpenAI has ever designated Critical for cybersecurity under its Preparedness Framework, so approved defenders in the Daybreak program got it first. ChatGPT Business and Pro followed on September 4 and Plus a few hours later, with the API, AWS, Azure and the free tier still outstanding. Artificial Analysis has since retired the Coding Agent Index, and on its own Terminal-Bench 4.0 run Astra now leads outright, 60% to Claude Fable 5.1's 55%, which is what moved the coding crown on this page to Astra. Full detail in our GPT-6 Astra breakdown.

Sep 2 Alibaba Qwen3.8-Max-0902 takes Arena's WebDev board off Claude Opus 5 by three points, at $2 / $6 New

Alibaba shipped Qwen3.8-Max-0902 on September 2, 2026, and the naming is the first thing to get right: this is a post-training refresh of the Qwen3.8-Max that launched on August 3, not a new version, running the same 2.4 trillion parameters and the same 1M-token context window. Qwen says it was further post-trained on coding and Cowork-style work. The reason it matters is the board it moved. Arena announced the same day that Qwen3.8-Max-0902 debuted at #1 overall on Code Arena: WebDev with 1691 points, three points above Claude Opus 5 (Max), seventeen above Kimi K3 (Max) and twenty-two above the previous Qwen3.8-Max, taking #1 in the Data & Analytics and Consumer Product categories as well. That is the first time this year an Anthropic model has been displaced at the top of a vote-based coding board. Pricing is $2 / $6 per 1M tokens with cache hits at $0.17 explicit and $0.25 implicit, which Arena scores as a blended $5 per million and the best position on its Pareto frontier. Read it with three caveats. Three points is inside the noise on a vote-based board, Artificial Analysis had published no Intelligence Index for the 0902 snapshot at the time and has since scored it 45.4 on v4.3, and this is a hosted model on QwenCloud with no consumer chat front-end: the open weights Alibaba promised for Qwen3.8-Max in August still have not shipped, so nothing here changes our open-weight pick. Arena says Agent Arena scores are still to come. Qwen3.8-Max-0902 has since slipped to fourth on WebDev, behind GPT-6 Astra, Claude Fable 5.1 and Claude Opus 5, and the coding category has moved to Astra. Full detail in our Qwen 3.8 breakdown.

Sep 2 Meta Muse Spark 1.3, Intelligence Index 57 to 61 at an unchanged $1.25 / $4.25, winning coding and losing every agent row New

Meta released Muse Spark 1.3 on September 2, 2026, its fourth Muse Spark model in five months, in Muse Code and the Meta Model API. Artificial Analysis scores it 61 on Intelligence Index v4.1.1, up from 57 for 1.2 and 53 for 1.1, which is eight points in two months and the fastest climb anyone on this page is managing. Pricing did not move at all: $1.25 / $4.25 per 1M tokens on a 1M-token context window, with the Contributor tier still at $0.10 / $0.20 in exchange for letting Meta train on your prompts and completions. Two things about the launch need reading carefully. On Meta's own scorecard the model wins every coding row, DeepSWE v1.1 at 75.4 against Claude Opus 5's 74.0 and SWEAtlas CodeBase QnA at 59.4, ties GPT-5.6 Sol on Terminal-Bench 2.1 at 88.8, and takes both MRCR long-context bands by a distance, 98.1 on the 512K to 1M band against 55.5 for 1.2. It also loses all six rows Meta groups under Agent, four to Opus 5 and two to GPT-5.6 Sol, and that half of the chart went almost unreported. The other catch is which model was measured: the benchmarked column is Muse Spark 1.3 (max), a limited preview for Meta's partners that scores 62, while the version anyone can call today is xhigh at 61, and Meta ran the 1.2 comparison column at xhigh rather than max. Meta shipped it to developers first, on Muse Code and the Meta Model API, and said a rollout to Facebook, Instagram and the Meta AI app would follow in the coming days, and the open-weights promise slipped again: August's pledge to release the Muse Spark 1.2 weights became an undated line about open weights coming soon. Full detail in our Muse Spark 1.3 breakdown.

Sep 2 Google Gemini 3.8 Flash, three points on the Intelligence Index at the same $0.75 / $3.75 New

Google made Gemini 3.8 Flash generally available on September 2, 2026, three weeks after 3.7 Flash and its third Flash release in 43 days. Every specification carries over unchanged: 1M input context, a 64K output limit, a March 2026 knowledge cutoff, the same modality list, and the same introductory $0.75 / $3.75 per 1M tokens that doubles to $1.50 / $7.50 on January 1, 2027. Only the scores moved, and they moved on every published row. DeepSWE v1.1 goes from 65.3% to 73.7%, Terminal-bench 2.1 from 85.8% to 89.4%, Terminal-bench 4.0 from 11.2% to 19.1% and BioMysteryBench Human Difficult from 43.5% to 56.5%. Artificial Analysis scores it 59 on Intelligence Index v4.1.1 against 56 for 3.7 Flash, and measures 304.6 output tokens per second against 279.4, with a slow 13.39-second time to first token that makes it a batch and agent model rather than an interactive one. Read the wins alongside the losses: Google's own comparison table gives 3.8 Flash 8 of 14 rows, and Claude Opus 5 still takes Terminal-bench 4.0 by 51.8% to 19.1%, OSWorld-2.0 by 75.4% to 59.0% and GDPVal-AA v2 by 1824 to 1545 on Elo. The DeepSWE figure republished across much of the launch coverage, 71.0%, is not the one in Google's PDF, which reads 73.7% against Opus 5's 74.0%. Developer access was live at announcement through the Gemini API, AI Studio, Antigravity and the Gemini Enterprise Agent Platform. Consumer access is reported for Google AI Pro and Ultra subscribers, but the Gemini Apps release notes carry no entry for it yet, and the free Gemini tier still runs 3.6 Flash. Full detail in our Gemini 3.8 Flash breakdown.

Sep 1 Anthropic Claude Fable 5.1 and Mythos 5.1, a new #1 on Artificial Analysis at 66 and a 75% cut to cache reads New

Anthropic released Claude Fable 5.1 and Claude Mythos 5.1 on September 1, 2026, and they are one model in two safeguard configurations rather than two models. Fable 5.1 is generally available; Mythos 5.1 relaxes the biology and cybersecurity restrictions and is handed out by invitation, with no published eligibility criteria. Artificial Analysis scores Fable 5.1 at 66 on its v4.1.1 Intelligence Index at max effort, the highest score it has ever measured, ahead of Claude Opus 5 at 63, Claude Fable 5 at 62, and GPT-5.6 Sol and Grok 4.6 at 61. It also took the then-current Agentic Index at 61 against Opus 5's 59, the GDPval-AA v2 professional-deliverables board at 1853 against 1824, and Humanity's Last Exam at 59.1% against Fable 5's 55.5%. The effort setting matters more than usual here: the same model reads 58 at low, 60 at medium, 62 at high and 65 at xhigh, and those settings span an eleven-fold difference in output tokens, from 13.1 million to 143.7 million across the evaluation suite. Pricing is unchanged at $10 / $50 per 1M tokens, but cache reads drop 75% to $0.25 from Fable 5's $1.00, which is where the saving actually sits on long agentic runs. On Anthropic's own figures it reaches 52.6% on Terminal-Bench-Science 0.1 against Opus 5's 29.0%, and the 212-page system card raised the company's own alignment risk assessment from very low to low. Two things to weigh before switching: Anthropic still recommends starting with Opus 5 for most workloads at half the price, and at launch no Arena board had votes for either model. Fable 5.1 has since been rated: it is third on Arena WebDev, second on image-to-WebDev and fifth on the text board. Full detail in our Claude Fable 5.1 and Mythos 5.1 breakdown.

Aug 28 Z.ai GLM 5.3 weights are out after the safety hold, but the licence is not MIT New

Z.ai published the GLM 5.3 weights on Hugging Face on August 28, 2026, two weeks after the API launch and on exactly the date its placeholder had carried, so the safety review it promised did happen and the hold is now closed rather than open-ended. What changed is the licence. GLM 5.3 Flash had shipped under plain MIT two days earlier; the flagship carries a bespoke document titled the GLM 5.3 License, and the model card's licence field reads glm-5.3 rather than mit. The grant itself tracks MIT closely, with rights to run, deploy, fine-tune, modify, distribute and sell, and with weights explicitly written into the definition of the software, so commercial use is open to almost everyone. One clause is genuinely new: a model-as-a-service operator whose revenue passes $10 billion over any twelve consecutive months has to pass a Z.ai security review before deploying commercially, and Z.ai reserves the right to decide that review's scope and method itself. Note also that Hugging Face counts the published checkpoint at 753B parameters against the 744B base Z.ai describes it sharing with GLM-5.2, which is the sort of gap a mechanical parameter count produces rather than evidence of a different model. This does not move our open-weight pick. GLM 5.3 Flash stays the model we would host, because 320B under plain MIT is a far easier thing to run and to clear legally than 753B under a bespoke licence. Full detail in our GLM 5.3 breakdown.

Aug 28 Tencent Tencent Hy4 Preview, 770B open weights under Apache 2.0 and the largest permissive model yet New

Tencent released and open-sourced Hy4 preview on August 28, 2026, the new flagship of its Hunyuan line and the largest model anyone has published under Apache 2.0. It runs 770 billion total parameters with 49 billion active per token on a mixture-of-experts design, takes a context window Tencent describes as exceeding 1M tokens, and ships with an FP8 build alongside the full-precision weights. API pricing is $0.834 per 1M input tokens, $2.501 output and $0.042 for cached input, which is inexpensive for the size. The licence is the real headline: Apache 2.0 is more permissive than Kimi K3's custom document and than the bespoke licence Z.ai attached to GLM 5.3 the same day. Treat the quality claim carefully, because the only evidence so far is the vendor's. Tencent ran an internal blind evaluation of 203 engineering tasks scored by 163 experts, where Hy4 preview takes 2.99 out of 4.00 against Kimi K3's 2.94 and GLM 5.3's 2.92. That is Tencent's own panel, the margins sit inside a tenth of a point, and no independent house had published an Intelligence Index or an Arena rating for it at the time. Arena has since placed it eleventh on WebDev at 1624, while Artificial Analysis still has no Intelligence Index for it, so we are listing it rather than ranking it. Tencent also says the model helped automate the optimisation of its own training pipeline and lifted its own inference throughput by a measured 31.8%. It is available as open weights and through WorkBuddy, CodeBuddy, Yuanbao and ima, plus Tencent Cloud TokenHub and OpenRouter, with free access on WorkBuddy and CodeBuddy for two weeks from launch. Full detail in our Tencent Hy4 Preview breakdown.

Aug 26 Z.ai GLM 5.3 Flash, Ox Alpha unmasked, and the first MIT weights Z.ai has shipped since GLM-5.2 New

Z.ai released GLM 5.3 Flash on August 26, 2026, and it is not the model the field was waiting for. The weights held back on August 14 belong to GLM 5.3; this is a separate model on a newly trained base, 320 billion parameters with 18 billion active, the first natively multimodal model in the GLM-5 line, and it went up on Hugging Face under MIT the same day. It is also the stealth model that had been serving traffic on OpenRouter as Ox Alpha, run during that preview entirely on Chinese AI chips, which describes how it was served rather than how it was trained. Artificial Analysis scores it 57 on the Intelligence Index v4.1.1, four points above GLM-5.2 at less than half the size, and places it on the intelligence-versus-cost Pareto frontier. On today's v4.3 index it reads 41.9 against GLM-5.2's 34.0, at $0.25 an index task. List price is $0.15 / $0.50 per 1M tokens; the 50% launch promotion ran to 24:00 on September 9 and has expired. DeepSeek's V4.1-Flash replacement on September 10 closed most of the gap: off-peak the two now tie at $0.15 on input, with GLM cheaper on output at $0.50 against $0.60, while at peak GLM undercuts it on both. Treat the coding and agentic figures with care: they come from Z.ai's own chart, which picks Claude Opus 4.8 and GPT-5.6 Terra as comparators rather than the current leaders, and it reads Terminal Bench 2.1 84.3, DeepSWE v1.1 63.4 against GLM-5.2's 46.2, and AutomationBench 48.8. The one number on that chart Artificial Analysis ran is GDPval-AA v2, where GLM 5.3 Flash takes 1773 Elo, the highest plotted. Arena has since rated it, seventeenth on WebDev and tenth on image-to-WebDev. This does move our open-weight pick: GLM 5.3 Flash replaces GLM-5.2 as the MIT model we would host. One caution on hardware, 18B active does not mean 18B to host, because mixture-of-experts routing needs all 320B weights resident, so this is a multi-GPU server rather than a workstation. Full detail in our GLM 5.3 Flash breakdown.

Aug 26 Alibaba Qwen3.8-Flash-Next, open weights and the first public look at the Qwen4 architecture New

Alibaba published Qwen3.8-Flash-Next on Hugging Face on August 26, 2026, and the point of it is the architecture rather than the scoreboard: Qwen says it is an early look at the design the Qwen4 family will use, released so the community can prepare for it. It is a 125-billion-parameter mixture-of-experts with only 6 billion parameters active per token, plus a separate 51-billion-parameter N-gram embedding, and it takes text, image and video in. Context is 262,144 tokens natively, extensible to 1M. Qwen's published figures include 91.7 on GPQA Diamond, 62.5 on SWE-Bench Pro, 58.7 on DeepSWE 1.1, 81.0 on SWE-Bench Multilingual and 84.5 on AndroidWorld vision, all vendor-reported and, at the time, with no independent rating at all. Artificial Analysis has since scored it 39.9 on Intelligence Index v4.3 and Arena places it ninth on WebDev at 1635. The licence needs reading before you build on it, because it is not Apache. The Qwen Community License 1.0 allows commercial use, modification and hosting, but it requires prominent display of the model name once a product passes 100 million monthly active users or $20 million in monthly revenue, and it requires a separate licence from Qwen to run a model-as-a-service or an AI work-assistant business on it. Internal use is exempt from that second condition.

Aug 21 OpenAI GPT-5.6 Sol cut by more than 20%, now under Claude Opus 5 on both sides New

OpenAI cut GPT-5.6 Sol from $5 / $30 to $4 / $20 per 1M tokens on August 21, 2026, the first cut to the flagship since the GPT-5.6 family launched in July. The rate is promotional and OpenAI says it runs about three months. It covers the API and credits on Codex and ChatGPT Work; Plus, Pro and Business subscription prices are unchanged. At $4 / $20 Sol now undercuts Claude Opus 5 at $5 / $25 on both input and output, the first time OpenAI's flagship has been the cheaper of the two.

Aug 16 DeepSeek DeepSeek raises API prices roughly fourfold and splits the rate card into peak and off-peak tiers New

DeepSeek moved both V4 models onto a new rate card at 16:00 UTC on August 16, 2026, and the direction is unusual for this field: prices went up, by roughly four times at peak. V4-Flash 0731 went from a flat $0.14 / $0.28 per 1M tokens to $0.22 / $0.66 off-peak and $0.44 / $1.32 at peak, and V4-Pro 0813 from $0.435 / $0.87 to $0.66 / $1.98 off-peak and $1.32 / $3.96 at peak, with cache-hit input at $0.007 and $0.022 off-peak respectively. Peak is 01:00 to 04:00 and 06:00 to 10:00 UTC, Monday to Friday, and every other hour is off-peak at exactly half the peak rate. DeepSeek framed the change as allocating capacity more reasonably. Nothing about either model changed, so this is purely an economics event, but it is a consequential one for this page: it cost DeepSeek V4-Flash the price-performance pick, which moved to GLM 5.3 Flash at $0.15 / $0.50 list. DeepSeek reversed most of the rise on September 10, retiring V4-Flash and replacing it with V4.1-Flash at $0.15 / $0.60 off-peak, so the two are level on input off-peak again. The weights are still MIT, so anyone self-hosting is unaffected. Full detail in our DeepSeek V4 breakdown.

Aug 14 Z.ai GLM 5.3, a large agentic jump from post-training alone, shipped without the open weights New

Z.ai released GLM 5.3 on August 14, 2026, and the notable thing is what did not change: it runs on the same 744-billion-parameter base as GLM-5.2, with no new pretrain. Every gain comes from scaled post-training, and the agentic ones are large. On Z.ai's own chart Terminal-Bench 3.0 goes from 4.6 to 28.3, DeepSWE v1.1 from 46.2 to 66.9, and CyberGym from 77.2% to 84.5%, which the company says edges Claude Mythos 5 at 83.8 and GPT-5.6 Sol at 83.6. Every one of those figures is the vendor's, with no independent audit yet, and the chart is less flattering elsewhere: on Z.ai's internal Code Bench, Claude Fable 5 still leads at 39.5% against GLM 5.3's 31.4%, and on ExploitBench the frontier leaders sit at 78.0% against 54.4%. The catch is the licence, not the score. GLM-5.2 shipped MIT weights within days; GLM 5.3 shipped with none, and Z.ai said they would follow after a safety evaluation of the model's vulnerability-discovery capability, roughly two weeks out. That resolved on August 28, 2026, when Z.ai published the weights on exactly the date the placeholder had carried, though under a bespoke GLM 5.3 License rather than the plain MIT that GLM-5.2 and GLM 5.3 Flash both got. Per-token access is now published at $1.40 / $4.40 per 1M tokens, alongside the GLM Coding Plan and ZCode. What arrived on August 26 instead was GLM 5.3 Flash, a different and smaller model that did ship MIT weights, and that is the release that moved our open-weight pick. Full detail in our GLM 5.3 breakdown.

Aug 13 Google Gemini 3.7 Flash, coding scores jump and the price halves to $0.75 / $3.75, until January New

Google shipped Gemini 3.7 Flash on August 13, 2026, 23 days after 3.6 Flash and its third Flash release in under two months, while Gemini 3.5 Pro remains unreleased. The coding gains are the point: DeepSWE v1.1 climbs from 49.0% to 65.3%, FrontierCode 1.1 Main from 34.4% to 43.6%, and AutomationBench nearly doubles from 17.0% to 30.4%. Artificial Analysis scores it 56 on its v4.1.1 Intelligence Index against 52 for 3.6 Flash, and ranks it first of 182 models on output speed at 385.1 tokens per second, which puts it on the intelligence-versus-time Pareto frontier. The context window and 64K output limit are unchanged from 3.6 Flash, so nothing was traded away. Two caveats carry more weight than the benchmarks. The $0.75 / $3.75 rate is introductory and expires on December 31, 2026, after which it doubles to $1.50 / $7.50, exactly what 3.6 Flash cost at launch; Google also moved 3.6 Flash onto the same rate, so the two now cost the same and the upgrade is a pure capability decision. And consumer access runs only through Spark, which needs Google AI Pro or Ultra and excludes the European Economic Area, the UK, Switzerland and Nigeria, so most European readers cannot reach it in the Gemini app at all. The free Gemini tier still gets 3.6 Flash. API access is unaffected everywhere. Full detail in our Gemini 3.7 Flash breakdown.

Aug 12 SpaceXAI Grok 4.6, post-training refresh takes Intelligence Index 56 to 61 at an unchanged $2 / $6 New

SpaceXAI released Grok 4.6 on August 12, 2026, and it is explicitly not a new foundation model: the company kept the Grok 4.5 base and bought the gains with a longer supplemental training run, regenerated fine-tuning trajectories and reinforcement learning inside agentic environments. No new architecture or parameter count was published. Artificial Analysis scores it at Intelligence Index 61, five points above Grok 4.5, which ties GPT-5.6 Sol for joint third behind Claude Opus 5 at 63 and Claude Fable 5 at 62 on the index as it stood that month; Claude Fable 5.1 has since gone in above all three at 66. The agentic figures are stronger than the composite: a GDPval-AA v2 Elo of 1753 behind only Opus 5, 88.4% on Terminal-Bench v2.1 and 50.7% on tau3-Banking. Efficiency was the argument at the time, at $0.84 per index task and roughly 53 turns and 0.5 billion input tokens on long-horizon work against about 103 turns and 2.0 billion for Opus 5. On Artificial Analysis's current v4.3 figures that argument has weakened: Grok 4.6 costs $1.86 an index task, more than Muse Spark 1.3 at $1.61 for a higher score. Pricing is unchanged at $2 / $6 per 1M tokens with cached input at $0.50, on the same 500K-token context window, but there is a trap worth knowing: cross 200,000 prompt tokens and SpaceXAI re-bills the entire request at $4 / $12, not just the overage. One warning on the benchmark coverage, because a lot of reporting has mangled it: Terminal-Bench v3.0 is a different and much harder test than v2.1, so Grok 4.6 scoring 26% on v3.0 and 88.4% on v2.1 are both true of the same model. It went live day one in the SpaceXAI API, Cursor and Grok Build, plus OpenRouter, Vercel and Cloudflare. Full detail in our Grok 4.6 benchmarks and pricing breakdown.

Aug 5 Meta Muse Spark 1.2 and Muse Code, Intelligence Index 51 to 57 at an unchanged $1.25 / $4.25, plus a terminal coding agent New

Meta shipped Muse Spark 1.2 on August 5, 2026 alongside Muse Code, its first terminal coding agent, which runs on macOS and Linux with no desktop app and no subscription. Artificial Analysis scores the model at Intelligence Index 57 at its highest reasoning setting on its current v4.1.1 index, up from 51 for 1.1 and 43 for the April release, and almost all of that gain is agentic rather than general knowledge; its GDPval-AA v2 Elo jumped from 1371 to 1631 while Terminal-Bench v2.1 moved only from 78% to 80%. Pricing is unchanged at $1.25 / $4.25 per 1M tokens on a 1M-token context window, and Meta added a Contributor tier at $0.10 / $0.20, roughly 92% cheaper, in exchange for letting Meta train on your prompts and completions. Availability widened from the US-only preview that 1.1 shipped under to expanded global access across Muse Code, the Meta Model API and OpenRouter. On Meta's own launch charts the model finishes second to Claude Opus 5 on all three coding benchmarks it published, at 82.9% against 86.7% on Terminal-Bench 2.1.

Aug 3 Alibaba Qwen 3.8 (Qwen3.8-Max), 2.4-trillion-parameter multimodal flagship now generally available at $2 / $6 New

Alibaba made Qwen3.8-Max generally available on August 3, 2026, two weeks after previewing it at WAIC Shanghai, and every figure that was missing at preview now has a number behind it. It runs 2.4 trillion total parameters with roughly 95 billion active per token on a sparse Mixture-of-Experts design, takes text, images and video, and confirms a 1M-token context window with up to 128K output tokens. API pricing is $2 / $6 per 1M tokens, flat across the entire context rather than tiered by prompt length, with cached reads at $0.25. The "second only to Fable 5" claim now splits cleanly in two: on Arena's vision board it is #2 at 1305, behind only Fable 5, but on the text board it is #5 at 1496, behind four Anthropic entries. Both Arena scores are still flagged preliminary, and the promised open weights have not shipped. We are keeping Qwen 3.8 out of the ranked picks until the weights and the licence actually land.

Pending Google Gemini 3.5 Pro, still unreleased, months behind schedule per Bloomberg Delayed

Gemini 3.5 Pro remains the biggest pending launch. Google announced it at I/O on May 19 alongside Gemini 3.5 Flash, but only Flash shipped, and the June target slipped. Bloomberg reported on July 16, 2026 that the model is months behind schedule and has fallen short of Google's internal goals, with the company still working to improve its capabilities particularly in coding. That is the whole sourced picture. Google has never publicly confirmed a launch date, and Google has not published final specs such as the context window or reasoning modes, so treat circulating figures as unconfirmed. Use Gemini 3.8 Flash in the meantime if you are working through the API, or Gemini 3.6 Flash if you are on the free Gemini app, which still runs it; we will move 3.5 Pro into the main ranking the moment it goes live.

Category pick, September 2026

Best AI for Writing

Claude Fable 5.1

Best writing overall

→

Kimi K3

Creative fiction

→

Claude Sonnet 5.5

Free everyday

The best AI for writing is Claude Fable 5.1, and the case for it is consistency rather than a single first place. It is inside the top seven of all three prose boards: second on EQ-Bench Creative Writing v3 at 2162.0, third on LiveBench Language at 89.5 and seventh on Arena's creative-writing board. No other model places that well across all three. Claude Opus 5.5 now leads Arena's creative-writing board at 1521 and its text board too, but it is seventh on EQ-Bench and eighth on LiveBench Language; GPT-6 Astra leads EQ-Bench at 2173.3 but sits 31st on Arena's creative-writing board; Claude Fable 5 tops LiveBench Language at 90.7 and is second on Arena's creative-writing board, but has slipped to eleventh on EQ-Bench. If your writing is a work deliverable rather than prose, the GDPval-AA v2.1 professional-deliverables board is a different question, and Claude Opus 5.5 leads it at 1846, with Claude Sonnet 5.5 two points behind at 1844 and Claude Fable 5.1 at 1735. Claude Sonnet 5.5 is the free-tier value pick, since anyone can use it on claude.ai, though of the three prose boards only LiveBench Language has rated it so far, at 83.4. Claude Fable 5 is the pick if you weight Arena's human votes above the harness scores; it costs $10 / $50 per 1M tokens and is permanently included in Claude Max and Team Premium at roughly 50% of regular usage limits. Translation is a separate question with a separate winner, which we work through in our guide to the best AI for translation.

Model	Best For	Strength	Weakness	Price (per 1M tokens)
Claude Fable 5.1	Best writing overall	#2 EQ-Bench Creative Writing (2162.0), #3 LiveBench Language (89.5) and #7 Arena creative writing, the best combined placing of any model	Leads none of the three outright; Claude Opus 5.5 now leads Arena's creative-writing board	$10 / $50
Claude Fable 5	Strong on the human-vote boards	#1 LiveBench Language (90.7), #2 Arena creative writing (1501.0), #3 Arena text (1503.8)	#11 EQ-Bench; priciest option here	$10 / $50
Kimi K3	Creative fiction and voice	#5 EQ-Bench Creative Writing (2082.3)	Only #25 on Arena creative writing	$3 / $15
Claude Sonnet 5.5	Free everyday writing	Free on claude.ai, 1M context; #2 GDPval-AA v2.1 (1844), two points behind Opus 5.5	83.4 LiveBench Language; not yet rated on EQ-Bench or Arena	$2 / $10
Claude Opus 5.5	Top of Arena's human-vote boards	#1 Arena creative writing (1521) and #1 Arena text (1509); #1 GDPval-AA v2.1 (1846)	#7 EQ-Bench and #8 LiveBench Language	$4 / $20
GPT-5.5	Fact-anchored business writing	Documented factual-reliability gains over GPT-5.4	Reasoning tiers now marked deprecated	$5 / $30
Gemini 3.8 Flash	Bulk drafts at scale	Intelligence Index 40.9, 308.4 tok/s output, 1M context	Agent-tuned rather than prose-tuned; intro price doubles January 1, 2027	$0.75 / $3.75

Runner-up and alternatives: Claude Fable 5 is the runner-up and the pick if you weight Arena's human votes, Claude Opus 5.5 is the pick if Arena's creative-writing board is the one you trust and the leader for professional deliverables, Kimi K3 is the pick for distinctive fiction though it now sits fifth on EQ-Bench, Claude Sonnet 5.5 is the value pick and the one to use if you are not paying, and Gemini 3.8 Flash is the pick for bulk drafting.

What changed this month

The writing crown stays with Claude Fable 5.1, but the ground under it moved again. Arena rated Claude Opus 5.5 within a week of its launch and put it first on both the text board, at 1508.6, and the creative-writing board, at 1521, which pushes Claude Fable 5 to third and second on those two boards. On the other two prose boards Opus 5.5 is only seventh on EQ-Bench and eighth on LiveBench Language, so Fable 5.1, at second, third and seventh, still has the best combined placing. It did slip from sixth to seventh on Arena's creative-writing board, so the claim on this card is now the top seven rather than the top six. EQ-Bench also re-scaled its Elo, which moved every figure in this block: GPT-6 Astra now leads at 2173.3, Fable 5.1 reads 2162.0 and Kimi K3 fell to fifth. Claude Sonnet 5.5, released September 28, replaces Sonnet 5 as the free pick. No creative-writing board has rated it yet, but it sits two points behind Opus 5.5 on the GDPval-AA v2.1 work-deliverables board at half Opus 5.5's token price. GPT-6.1 Sol, released September 29, entered LiveBench Language at 90.13, second behind Claude Fable 5 and ahead of Fable 5.1, which drops to third there; neither EQ-Bench nor Arena has rated it yet, so it cannot yet match Fable 5.1's placing across all three.

Category pick, September 2026

Best AI for Chat & Daily Assistant

GPT-5.6

ChatGPT's default

→

Claude Opus 5.5

Highest-rated

→

Claude Sonnet 5.5

Free Claude

The best AI for everyday chat is GPT-5.6, and the honest reason is reach rather than board position. It is the model ChatGPT serves by default to the largest user base in the category, which makes it the best assistant most people can actually open. On raw human preference it is not the leader: GPT-5.6 Sol sits #19 on Arena's text leaderboard at 1483.2, where Claude Opus 5.5 leads at 1508.6. Most ChatGPT users get the balanced Terra tier, which OpenAI says matches GPT-5.5 and, since the July 30, 2026 price cut, costs 60% less than it. It is available inside ChatGPT (free with limits, Plus at $20/month, Pro at $100/month), through the API (the 5.6 tiers: Luna $0.20 / $1.20, Terra $2 / $12, Sol $4 / $20 per 1M tokens), and bundled inside Fello AI alongside Claude, Gemini, Grok, and DeepSeek. One caveat: OpenAI's system card and the evaluator METR flagged elevated "scheming" behaviour in Sol, so GPT-5.5 Instant stays the safer pick for hallucination-sensitive work.

Model	Best For	Strength	Weakness	Price
GPT-5.6	Everyday chat, ChatGPT's default	The assistant most people can open; Terra matches GPT-5.5 at ~half cost	#19 on Arena text; scheming flagged by METR	Free / $20/mo Plus; API $0.20 / $1.20 to $4 / $20
Claude Opus 5.5	Highest-rated conversation	#1 on Arena text overall (1508.6) and #1 on the Intelligence Index (57.6)	Not on the Free plan; usage-limited on Pro	$20/mo Pro, $4 / $20 API
Claude Sonnet 5.5	Free Claude chat	Free on claude.ai; #2 on the Intelligence Index (56.0), behind only Claude Opus 5.5	Not yet rated on Arena text; runs at Medium effort by default in the apps	Free / $20/mo Pro, $2 / $10 API
GPT-5.5 Instant	Hallucination-sensitive daily work	52.5% fewer hallucinated claims vs 5.3 Instant	Reasoning tiers now marked deprecated	$20/mo Plus; API $5 / $30
Gemini 3.6 Flash	Fast, free, multimodal	Still the model the free Gemini tier serves, 1M context, #20 on Arena text	Weaker on hardest reasoning; 3.8 Flash outscores it at the same price	Free / $0.75 / $3.75 API
Fello AI	All the top models, one app	ChatGPT + Claude + Gemini + Grok + DeepSeek and more on Mac, iPhone and iPad	Routed via app, not direct	$9.99/mo

Runner-up and alternatives: Claude Opus 5.5 is the runner-up and the preference leader on Arena's text board, Claude Sonnet 5.5 is the pick for free Claude chat, Gemini 3.6 Flash is the runner-up for fast and free, and Grok 4.6 is the niche pick for live-news days. Fello AI is the natural pick if you want the top models in one Mac and iOS app for $9.99/month instead of juggling subscriptions.

What changed this month

GPT-5.6 keeps the chat pick on reach, but the preference leader behind it changed. Arena has now rated Claude Opus 5.5, and it went straight to #1 on the text board at 1508.6, ahead of Claude Opus 4.6 and of Claude Fable 5, which led that board last month and is now third. GPT-5.6 Sol has slipped to #19 at 1483.2. So Opus 5.5 takes the second podium place from Fable 5, and Claude Sonnet 5.5, released September 28 and free on claude.ai, takes the third from Claude Opus 5, which Anthropic now lists as a legacy model. None of that moves the crown, which is about which assistant the most people can actually open. GPT-6 Astra reached ChatGPT Business, Pro and Plus in early September but OpenAI has announced nothing for the free tier, and OpenAI put GPT-6 Sol and GPT-6 Luna in ChatGPT Work and Codex on September 22, saying plainly that they are not in Chat yet, so the window most people type into still runs GPT-5.6. GPT-6.1 Sol took the same route on September 29: ChatGPT Work and Codex, not Chat.

Category pick, September 2026

Best AI for Images

ChatGPT Images 2.5

Text in images

→

MAI-Image-2.6

#1 AA image editing

→

Muse Image

Meta AI ecosystem

The best AI for image generation is ChatGPT Images 2.5, which OpenAI made the default on every ChatGPT tier on September 8 and which now leads all four image boards we track. Its two models finish first and second on each: GPT-Image-2.5 Sunburst, built for precision, leads Arena text-to-image at 1423.2 and Arena image editing at 1526.0, and it leads Artificial Analysis's quality Elo at 1196.8 and its image-editing board at 1221.3, with the faster GPT-Image-2.5 Flare second on all four. That is a change from last month, when Artificial Analysis had rated neither model and its boards disagreed with Arena's. Read the Sunburst-to-Flare gap as provisional even so, because those Arena placements rest on a few thousand votes each against GPT Image 2's 83,000 and 259,000, and on text-to-image the two share a rank band. MAI-Image-2.6 is the runner-up, third on Artificial Analysis's image-editing board at 1191.6 and fourth on Arena's text-to-image board at 1334.0, with Meta's Muse Image third. Reve 2.1 keeps its case for layout, typography and native 4K but has left the podium: it is sixth on Arena text-to-image and fourteenth on Arena image editing. Our full breakdown of what shipped is in ChatGPT Images 2.5.

Model	Best For	Strength	Weakness	Price
ChatGPT Images 2.5	Images with readable text	#1 and #2 on all four boards: Arena text-to-image (1423.2 / 1401.3), Arena editing (1526.0 / 1481.8), Artificial Analysis quality (1196.8 / 1190.4) and editing (1221.3 / 1207.0)	Arena placements rest on a few thousand votes each	Default on every ChatGPT tier
MAI-Image-2.6	Editing an image you already have	#3 on Artificial Analysis image editing (1191.6) and #4 on Arena text-to-image (1334.0)	Public preview, not in Copilot or Bing yet; lost the Artificial Analysis editing lead to ChatGPT Images 2.5	MAI Playground / Microsoft Foundry
Muse Image	Image editing, Meta ecosystem	#5 on Artificial Analysis image editing (1171.3) and #7 on its quality Elo	New, thin tooling around it	Meta AI app
Nano Banana 2 (Gemini 3.1 Flash Image)	Photoreal portraits and products	#4 on Artificial Analysis text-to-image, ahead of Nano Banana Pro on Arena's	Nano Banana Pro is ahead on Arena image editing	Gemini app / AI Studio
Seedream 5.0 Pro	Multilingual text + region editing	10+ languages incl. Arabic RTL, lasso and layer editing	Around tenth on the independent boards; copyright cloud	BytePlus / Magnific
Midjourney v8	Stylized art, illustration	Aesthetic baseline most artists prefer	Weaker on text in image	$10-$120/mo
Grok Imagine	NSFW / Spicy Mode	Most permissive guardrails; Arena rates its Image 2.0 model #5 on text-to-image and #4 on image editing, and Artificial Analysis now has it #4 on quality Elo	Nothing in the release record shows Image 2.0 shipping as a product	$30/mo SuperGrok

Runner-up and alternatives: MAI-Image-2.6 is the runner-up overall and the pick for editing an image you already have, Muse Image is third and the pick inside Meta's apps, and Nano Banana 2 is still the photoreal pick. Reve 2.1 remains the pick for layout, typography and native 4K even after leaving the podium. Grok Imagine is still the only frontier model that allows Spicy Mode adult content, and Arena now rates its Image 2.0 model fifth on text-to-image and fourth on image editing.

What changed this month

ChatGPT Images 2.5 keeps the image crown and its case got stronger. Last month Artificial Analysis had rated neither of its two API models, so the boards we track disagreed: Arena had Sunburst and Flare first and second while Artificial Analysis still had GPT Image 2 on top of text-to-image and Microsoft's MAI-Image-2.6 on top of image editing. Artificial Analysis has now rated both, and they take first and second there too, 1196.8 and 1190.4 on the quality Elo and 1221.3 and 1207.0 on image editing. That makes it #1 and #2 on all four boards. MAI-Image-2.6 stays the runner-up but slips to third on the editing board it used to lead, and Grok Imagine Image 2.0 is now fourth on Artificial Analysis's quality Elo and fourth on Arena's editing board, which is a better placing than the podium's third; we still report it rather than crown it, because nothing in the release record shows it shipping as a product.

Category pick, September 2026

Best AI for Video

Gemini Omni 1.1 Flash

#1 Arena text-to-video

→

MiniMax H3

2K + native audio

→

Dreamina Seedance 2.0

Closest challenger

The best AI for video generation is Gemini Omni 1.1 Flash, which Google shipped on August 27 and which leads Arena's text-to-video board at 1516, just ahead of its own predecessor Gemini Omni Flash at 1513. Read that as a tie: the two sit inside each other's confidence intervals and Arena itself gives 1.1 Flash a rank band of one to four. What breaks it is specification, where 1.1 Flash is clearly the more capable model: scene extension in 10-second increments to a cumulative 40 seconds, first and last frame control, video references, 360p drafts and 1080p or 4K output, priced at roughly $0.03 per second at 360p, $0.10 at 720p, $0.15 at 1080p and $0.30 at 4K. Gemini Omni Flash remains the cheaper option at about $0.10 a second but caps at 10-second clips. Two limits worth knowing. Artificial Analysis has not rated 1.1 Flash at all; on its text-to-video board the older Omni Flash leads at 1512.8 from Alibaba's Wan 3.0 at 1431.4 and MiniMax H3 at 1407.7. And on Arena's image-to-video board 1.1 Flash is second at 1488.5 behind MiniMax H3 at 1495.4, so the crown rests on text-to-video and on specification, not on a sweep. If you need longer takes inside Google's tooling, Veo 3.1 still runs in the Gemini app, AI Studio and Vertex AI with native audio and 1080p output. MiniMax H3 (July 31) remains the specification contender, with 2K clips of 4 to 15 seconds and native stereo audio from $0.13 a second.

Model	Best For	Strength	Weakness	Price
Gemini Omni 1.1 Flash	Best AI video overall	#1 Arena text-to-video (1516); 40-second scene extension, 4K output	Unrated by Artificial Analysis; second to MiniMax H3 on Arena image-to-video	$0.03-$0.30/sec; Gemini API / AI Studio / Flow
Gemini Omni Flash	Cheaper 10-second clips	#2 Arena text-to-video (1511), #2 AA video (1238); conversational editing	Caps at 10-second generations	~$0.10/sec; Gemini app / AI Studio
Wan 3.0	Video from documents and slides	#1 on Artificial Analysis's video board (1239); 2-30s, up to 1080p, Omni-Reference input	API-only, no published weights; #3 on Arena	$0.05-$0.20/sec
MiniMax H3	2K clips with native audio	4-15s at 2K, native stereo audio; #8 Arena text-to-video (1462)	#4 on AA video (1227); the open weights exclude the 2K upscaler and need an application in the US, EU, UK and South Korea	$0.13/sec at 2K
Dreamina Seedance 2.5	ByteDance challenger	#6 Arena text-to-video (1482); leads image-to-video with audio	ByteDance ecosystem, limited Western access	Dreamina / BytePlus
Muse Video	Meta ecosystem video	#3 text-to-video at 1459	Newest of the group, thin tooling	Meta AI app
Veo 3.1	Longer production clips	Native audio, 1080p, strong physics consistency	#6 on Arena video, not the quality leader	Google AI Pro / Ultra
Kling 3.0 / 3.0 Turbo	Fast iteration at lower cost	Native 4K, 60fps, 15-second clips; Turbo shipped June 17	Outside the top 16 on Arena text-to-video	From $10/mo
Luma Ray 3	Photoreal scenes	Strong realism for landscapes	Smaller community	Free / from $9.99/mo

Runner-up and alternatives: MiniMax H3 is the runner-up and now leads Arena's image-to-video board at 1495.4, ahead of Gemini Omni 1.1 Flash at 1488.5. Dreamina Seedance 2.0 is third and still the pick for image-to-video with audio. Veo 3.1 is the pick when 10 seconds is not enough, and Alibaba's Wan 3.0 is the one to use when the source material is a document, a spreadsheet or a slide deck. OpenAI retired the Sora 2 consumer app on April 26, 2026 and only the developer API remains, through September 24, 2026. We track the retirement dates for older models across every provider in a separate running list.

What changed this month

Gemini Omni 1.1 Flash keeps the video crown, but the picture around it is messier than last month's entry implied. It still leads Arena's text-to-video board, 1516 to Gemini Omni Flash's 1513, and Arena gives it a rank band of one to four, so that remains a tie broken on capability. Two things we reported in August have since changed. Artificial Analysis re-anchored its video Elo and the old 1239 / 1238 / 1235 trio no longer exists: Gemini Omni Flash now leads that board at 1512.8 from Wan 3.0 at 1431.4 and MiniMax H3 at 1407.7, and Artificial Analysis has still not rated 1.1 Flash at all. And MiniMax H3 has taken the top of Arena's image-to-video board at 1495.4, pushing 1.1 Flash to second at 1488.5, which is the first board this crown has lost.

Category pick, September 2026

Best AI for Coding

Claude Opus 5.5

#1 Arena WebDev

→

Claude Sonnet 5.5

Value pick, #1 Terminal-Bench

→

GPT-6 Astra

#1 image-to-WebDev

The best AI for coding is Claude Opus 5.5, which Anthropic shipped on September 22, but it no longer leads every board, and the model that took two of them is its own cheaper sibling. Claude Sonnet 5.5, released September 28, leads Artificial Analysis's own Terminal-Bench 4.0 run at 63.6% against Opus 5.5's 59.6% and LiveBench Coding at 91.4 against 89.3. Opus 5.5 keeps the crown on the rest of the evidence. Arena has now rated it and put it first on WebDev at 1820, ahead of GPT-6 Astra at 1792. It scores 71.7 on LiveBench Agentic Coding, second only to DeepSeek V4.1-Flash, where Sonnet 5.5 manages 56.3, and on Anthropic's own harnesses it beats Sonnet 5.5 on FrontierCode v1.1, 54.4% to 52.1%, and on CursorBench 4.0, 57.8% to 55.5%. Price does not settle it the way the token rates suggest. Sonnet 5.5 is $2 / $10 per 1M tokens against Opus 5.5's $4 / $20, but at max effort it spends so many tokens that an Intelligence Index task costs $7.60 against Opus 5.5's $5.98. Its value is at lower effort: at xhigh it scores 51.9 on the index for $2.74 a task and at high 46.7 for $1.08, and Anthropic's own Terminal-Bench chart puts it at Medium, the default in Claude Code, far above Sonnet 5's best score. GPT-6 Astra keeps Arena's image-to-WebDev board at 1733 and matches Opus 5.5's 59.6% on Terminal-Bench 4.0 at xhigh, at $10 / $50. Claude Fable 5.1 is third on WebDev and second on image-to-WebDev. Among open weights, Xiaomi's MiMo-V2.6-Pro takes 34.8% on Terminal-Bench 4.0 under plain MIT at $0.13 an index task, GLM-5.3 is still the strongest open model on that harness at 41.9%, and Kimi K3 is the highest-placed open model on Arena WebDev, tenth at 1659, on only 12.6% of the harness. Artificial Analysis has retired both its Coding Index and its Coding Agent Index, so the two composite coding rankings this page used to quote no longer exist. GPT-6.1 Sol, released September 29 at $2 / $10, is the cheap OpenAI option: 51.8 on the index and 56.1% on Terminal-Bench 4.0 for $0.72 an index task, though LiveBench puts it well outside the top ten on Coding, at 80.7, and Arena has not ranked it yet. Arena has now ranked Claude Sonnet 5.5, fourth on WebDev at 1699.

Model	Best For	Strength	Weakness	Price (per 1M tokens)
Claude Opus 5.5	Best coding overall	#1 Arena WebDev (1820) and #1 Intelligence Index (57.6, v4.3.2); 71.7 LiveBench Agentic Coding, 15 points above Sonnet 5.5	Second to Sonnet 5.5 on Terminal-Bench 4.0 (59.6% to 63.6%) and LiveBench Coding (89.3 to 91.4)	$4 / $20
Claude Sonnet 5.5	Value pick for coding	#1 Terminal-Bench 4.0 (63.6%) and #1 LiveBench Coding (91.4); 51.9 on the index for $2.74 a task at xhigh	56.3 LiveBench Agentic Coding; at max effort $7.60 an index task, more than Opus 5.5; only #4 on Arena WebDev (1699)	$2 / $10
GPT-6 Astra	Top of Arena image-to-WebDev	#1 Arena image-to-WebDev (1733), #2 WebDev (1792); 59.6% Terminal-Bench 4.0 at xhigh	2.5x Opus 5.5's price; behind Sonnet 5.5 on Terminal-Bench and well behind both Claude models on LiveBench Coding	$10 / $50
Claude Fable 5.1	Highest composite score before Opus 5.5	#3 Arena WebDev and #2 image-to-WebDev; Intelligence Index 53.4; AA-Briefcase 1678	52.0% Terminal-Bench 4.0; 2.5x the price of Opus 5.5	$10 / $50
Claude Opus 5	Superseded by Opus 5.5	49.0% Terminal-Bench 4.0 and #5 Arena WebDev and #3 image-to-WebDev	Costs more per token than Opus 5.5 and scores seven points lower	$5 / $25
Claude Fable 5	Hardest long-horizon agentic work	Intelligence Index 49.6; 42.4% Terminal-Bench 4.0; 1M context	The priciest per index task on this table at $8.75; #6 on Arena image-to-WebDev	$10 / $50
Qwen3.8-Max-0902	Brief holder of Arena's WebDev board	#8 Code Arena: WebDev (1671); Intelligence Index 45.4; 2.4T parameters, 1M context	QwenCloud API only with no consumer front-end; lost the WebDev lead within days	$2 / $6
Kimi K3	Highest-placed open model on Arena	#10 Arena WebDev (1659), the best open placement on that board; 2.8T/104B active	Only 12.6% on Terminal-Bench 4.0; 1.56 TB to self-host	$3 / $15
GPT-5.6 Sol	OpenAI's cheaper flagship	Intelligence Index 47.0; 39.9% Terminal-Bench 4.0	Superseded by GPT-6 Sol, which scores higher at half the price; eval-gaming flagged by METR	$4 / $20
GPT-6 Sol	Cheap GPT-6 coding in Codex	Intelligence Index 47.5 (v4.3.2) and 43.9% Terminal-Bench 4.0, both above GPT-5.6 Sol, at half its price	#6 Arena WebDev (1692); in Codex and ChatGPT Work only, not in Chat	$2 / $10
GPT-6.1 Sol	Cheapest top-tier model per task	Intelligence Index 51.8 (v4.3.2) and 56.1% Terminal-Bench 4.0 at $0.72 an index task, with cached input at $0.10	80.7 on LiveBench Coding, well outside the top ten; not yet rated on Arena; in Codex and ChatGPT Work only, not in Chat	$2 / $10
Muse Spark 1.3	Cheap agentic coding	Intelligence Index 48.1 at $1.25 / $4.25, and $1.60 per index task	33.3% on Terminal-Bench 4.0; the coding wins come from Meta's own chart, measured on a partners-only max variant	$1.25 / $4.25
Grok 4.7	Cheap value coder	46.3% CursorBench 4.0 and 71.0% DeepSWE v1.1 on SpaceXAI's own figures; Intelligence Index 46.3	24.7% on Terminal-Bench 4.0; prompts of 200K tokens and up re-bill the whole request at double	$2 / $6
Gemini 3.8 Flash	Agent coding at scale	Intelligence Index 40.9; 308.4 output tokens per second	19.7% on Terminal-Bench 4.0; intro price doubles January 1, 2027	$0.75 / $3.75
GLM 5.3 Flash	Cheapest serious coder you can host	32.8% Terminal-Bench 4.0 at $0.25 an index task; MIT, 320B/18B active	GLM-5.3 reaches 41.9% on the same harness; no consumer product	Open weights (MIT)
MiMo-V2.6-Pro	Highest open Intelligence Index	46.3 on v4.3.2 and 34.8% Terminal-Bench 4.0 at $0.13 an index task; MIT, 1.02T/42B active	Only #21 on Arena WebDev (1618), eleven places behind Kimi K3	$0.435 / $0.87

Runner-up and alternatives: Claude Sonnet 5.5 is the runner-up and the value pick, leading Terminal-Bench 4.0 and LiveBench Coding at half Opus 5.5's token price as long as you keep it below max effort, GPT-6 Astra is third and still tops Arena's image-to-WebDev board, Claude Fable 5.1 is third on WebDev, and Claude Opus 5 is now superseded by a model that is both cheaper and better. Among open weights, Xiaomi's MiMo-V2.6-Pro is the new leader on score and on cost, GLM-5.3 is the strongest on Artificial Analysis's Terminal-Bench 4.0, Kimi K3 is the highest-placed open model on Arena's WebDev board, and GLM 5.3 Flash is the pick for teams hosting their own, at $0.25 an index task under plain MIT. Inside IDEs, Cursor with Claude is still the most popular pairing and Claude Code is the natural pick if you live in the terminal.

What changed this month

Claude Opus 5.5 keeps the coding crown, but it is now a split with its own cheaper sibling rather than a sweep. Claude Sonnet 5.5, released September 28, took two of the boards this block rests on: Artificial Analysis's Terminal-Bench 4.0, 63.6% against Opus 5.5's 59.6%, and LiveBench Coding, 91.4 against 89.3. The crown stays with Opus 5.5 because it wins more of the evidence. Arena has now rated it and put it first on WebDev at 1820, ahead of GPT-6 Astra at 1792. It leads Sonnet 5.5 by 15 points on LiveBench Agentic Coding and by about two points on Anthropic's own FrontierCode and CursorBench runs, and at max effort it is the cheaper of the two per index task, $5.98 against $7.60. Sonnet 5.5 takes second place on the podium as the value pick and GPT-6 Astra drops to third, keeping only Arena's image-to-WebDev board, so Claude Fable 5.1 leaves the podium. One correction to the September 22 entry: it said Opus 5.5 led LiveBench Agentic Coding, but DeepSeek V4.1-Flash leads that board at 77.3, with Opus 5.5 second at 71.7. Arena's WebDev order moved further down too: Claude Sonnet 5.5 has since entered fourth at 1699, which pushes GPT-6 Sol to sixth, Kimi K3 to tenth and MiMo-V2.6-Pro to twenty-first, with Qwen3.8-Max-0902 still eighth. GPT-6.1 Sol, released September 29, does not challenge the crown: it scores 51.8 on the index and 56.1% on Terminal-Bench 4.0, both below Sonnet 5.5, and 80.7 on LiveBench Coding, but at $0.72 an index task it is the cheapest route to a top-tier score.

Category pick, September 2026

Best AI for Creativity

Grok 4.7

Fewest guardrails

→

Claude Fable 5

Highest-quality prose

→

Kimi K3

Fiction and voice

The best AI for unfiltered, on-trend creative work is Grok 4.7, and we want to be exact about why. This pick is about the product line, not the prose quality. Grok carries the fewest content restrictions of any frontier model and the only native real-time X integration, which makes it the one model that will engage with edgy, topical or deliberately provocative briefs that the others decline. Read the availability carefully: SpaceXAI shipped Grok 4.7 on September 21 to the Grok API, Cursor, Grok Build, third-party harnesses and cloud routers, and its announcement says nothing about the consumer app, which still serves Grok 4.6 to SuperGrok and X Premium+ subscribers at $30/month. It is not the best writer, though the picture is more mixed than it was. EQ-Bench has now rated Grok 4.7 and puts it eighth at 2006.7, inside its top ten and far above Grok 4.5 in 52nd, but Arena's human voters put 4.7 only 72nd on creative writing, behind Grok 4.6 in 51st and the older Grok 4.20-beta1 in 21st. If you are picking on output quality alone, Claude Opus 5.5 now leads Arena's creative-writing board with Claude Fable 5 second, while EQ-Bench puts GPT-6 Astra first at 2173.3, Claude Fable 5.1 second and Kimi K3 fifth. Choose Grok for what it will let you make, not for how well it writes.

Model	Best For	Strength	Weakness	Price
Grok 4.7	Unfiltered, opinionated, on-trend	Fewest content restrictions, native real-time X grounding	#72 on Arena creative writing, though #8 on EQ-Bench (2006.7)	$2 / $6 API; $30/mo SuperGrok app, still on 4.6
Claude Fable 5	Highest-quality creative prose	#2 Arena creative writing, #1 LiveBench Language	Cautious guardrails on edgy briefs	$10 / $50 API
Kimi K3	Fiction and distinctive voice	#5 EQ-Bench Creative Writing at 2082.3	Only #25 on Arena creative writing	$3 / $15
Claude Opus 5	Long-form structured creativity	Holds long threads and self-edits; Intelligence Index 50.8, behind Claude Opus 5.5, Claude Fable 5.1 and GPT-6 Astra	Most cautious of the group	$20/mo Pro; $5 / $25 API
Gemini 3.1 Pro	Multimodal creative	Strong text, image and video chain	Quotas inside the Gemini app	Free / $2.00-$4.00 API in
Grok Imagine (Spicy Mode)	NSFW / adult creative	Most permissive image generation	Niche use case	$30/mo SuperGrok

Runner-up and alternatives: Claude Fable 5 is the runner-up and the right pick if quality matters more than freedom, Kimi K3 is the pick for fiction, and Claude Opus 5 is the pick for creative projects that run across many turns. For adult creative work, Grok Imagine Spicy Mode is still the only frontier-grade option.

What changed this month

Grok keeps this pick on Grok 4.7, and nothing about the product's permissiveness or its live X access changed, which is what the pick rests on. What changed is that both creative-writing boards have now rated 4.7, and they disagree sharply. EQ-Bench puts it eighth at 2006.7, a large step up from Grok 4.5 in 52nd, while Arena's human voters put it 72nd, below Grok 4.6 in 51st. That corrects two lines in the September 22 entry, which said no creative-writing board had rated 4.7 and that Grok 4.6 sat 32nd on Arena; it has since slipped to 51st. The split surface still matters for this category: 4.7 is on the API, Cursor and Grok Build, while the $30/month Grok app that the permissiveness argument is really about still serves 4.6. For better writing, Claude Opus 5.5 now leads Arena's creative-writing board and Claude Fable 5 is second.

Category pick, September 2026

Best AI for Accuracy & Research

Claude Fable 5.1

Highest factual accuracy

→

Gemini 3.1 Pro

Best value, $0.52/task

→

GPT-5.6 Sol

Novel reasoning

The best AI for accuracy and research is Claude Fable 5.1, which holds the highest factual accuracy Artificial Analysis has measured, 67% on AA-Omniscience. That is the exact column this category is decided on, and it survived a hard test on September 22: Claude Opus 5.5 beat Fable 5.1 on almost everything else, including Humanity's Last Exam at 61.4% to 59.1% and the hallucination-penalised AA-Omniscience index at 46.4 to 43.5, but on raw factual accuracy it reads 66% to Fable 5.1's 67%. Treat that one-point gap as a tie on its own and read the two together: Opus 5.5 is the better bet when a wrong answer is worse than no answer, Fable 5.1 when you want the most facts right. The value pick is Gemini 3.1 Pro, and it is still what we would reach for when cost matters. Its strongest result is on ARC Prize's ARC-AGI-1, where it scores 98% and ties the human panel, and it does that at $0.52 per task, though four models have since gone past the panel at 98.5%, among them Claude Fable 5, GPT-6 Astra and Claude Opus 5.5. It pairs that with native Google Search grounding, which is what you want when the answer has to be current rather than merely plausible. It also scores 94.1% on GPQA Diamond, where GPT-5.6 Sol now matches it rather than trailing it, and 44.4% on Humanity's Last Exam, and scores 46.44 on Scale SEAL's HLE board, which Claude Fable 5 now leads at 55.5%. Two honest caveats. On grounded search specifically, Arena's search leaderboard is led by OpenAI rather than by Google or Anthropic: GPT-5.6 Sol tops it at 1257, with Claude Fable 5 fifth and Gemini 3.1 Pro grounding ninth. And on novel reasoning the crown sits elsewhere entirely, with GPT-6 Astra leading both ARC-AGI-2 and ARC-AGI-3. We did not move this crown to Claude Opus 5 because Artificial Analysis measures its hallucination rate at 50% and places it well below both Fable models on AA-Omniscience.

Model	Best For	Key Benchmark	Weakness	Price
Claude Fable 5.1	Highest measured factual accuracy	67% factual accuracy on AA-Omniscience, the highest Artificial Analysis has measured, one point above Claude Opus 5.5	No cheap tier; unrated on Arena's search leaderboard; Opus 5.5 leads Humanity's Last Exam and the AA-Omniscience index	$10 / $50
Gemini 3.1 Pro	Best value, cheap factual work	98% ARC-AGI-1 (ties human panel) at $0.52/task, 94.1% GPQA (tied by Sol)	ARC-AGI-2 77.1% now ranks ~14th; #9 on Arena search	$2.00-$4.00 / $12.00-$18.00 (tiered)
Claude Fable 5	Grounded search	#5 on Arena's search leaderboard, behind GPT-5.6 Sol	No single cheap tier	$10 / $50 (Fable 5)
GPT-5.6 Sol	Novel reasoning	#3 ARC-AGI-2 at 92.5%, behind GPT-6 Astra (95%) and Claude Opus 5.5 (93.3%)	Scheming flagged by METR	$4 / $20
Claude Opus 5	Hardest unseen problems	#1 ARC-AGI-3 at 30%, ~3.75x the next model (ARC Prize)	Artificial Analysis measures a 50% hallucination rate	$5 / $25
Qwen 3.7 Max	Frontier accuracy at value pricing	92.4 GPQA Diamond, 200 free requests/day	API-only, no chat front-end	$1.25 / $3.75 promo; $2.50 / $7.50 list
Claude Opus 4.6	Honesty under pressure	#1 on Scale SEAL's MASK board at 96.28; Anthropic holds the top 5	Superseded as a flagship	Legacy Anthropic model

Runner-up and alternatives: Gemini 3.1 Pro is the runner-up and the value pick at $0.52 a task with Google Search grounding, GPT-5.6 Sol is the runner-up for novel reasoning, Claude Opus 4.6 sweeps the honesty-under-pressure board, and Qwen 3.7 Max is the value pick at the frontier.

What changed this month

The accuracy crown stays with Claude Fable 5.1, and it was tested hard. Claude Opus 5.5 arrived on September 22 and took Humanity's Last Exam at 61.4% against 59.1% and the hallucination-penalised AA-Omniscience index at 46.4 against 43.5, but this category is decided on the raw factual-accuracy column and there Fable 5.1 still reads 67% against Opus 5.5's 66%. That is a one-point gap, so the block now says plainly that the two split rather than pretending Fable 5.1 sweeps: use Opus 5.5 when a wrong answer is worse than no answer, Fable 5.1 when you want the most facts right. Two corrections to last month's entry as well. It said GPT-5.6 Sol led ARC-AGI-2 and Claude Opus 5 led ARC-AGI-3; GPT-6 Astra has led both since its September 3 launch, as the problem-solving block already said, and that contradiction is gone. And ARC-AGI-1 has moved past the human panel: Claude Fable 5 and GPT-6 Astra both score 98.5% where the 2025 panel scores 98%, so Gemini 3.1 Pro's 98% at $0.52 a task is still true and still the value argument, but it is no longer the top of that board. GPT-6 Sol is the other accuracy move of the month, and it happens below the crown: Artificial Analysis measures its hallucination rate at 60% against GPT-5.6 Sol's 92%, while raw factual accuracy edges down from 59% to 55%, which lifts its AA-Omniscience index from 22.0 to 27.1. Claude Sonnet 5.5, released September 28, does not challenge this crown: Artificial Analysis puts its raw factual accuracy at 54% and its AA-Omniscience index at 32.3, well below both leaders. GPT-6.1 Sol, released September 29, lands below both leaders: Artificial Analysis measures its raw factual accuracy at 62% and its AA-Omniscience index at 41.5, with the hallucination rate down to 54% from GPT-6 Sol's 60%.

Category pick, September 2026

Best AI for Problem Solving

GPT-6 Astra

Reasoning leader

→

Claude Opus 5

Agentic chains

→

GPT-6.1 Sol

Value pick, $2 / $10

The best AI for hard problem solving is GPT-6 Astra, which took the reasoning boards off GPT-5.6 Sol within days of launching. It leads LiveBench Reasoning at 92.65 and ARC-AGI-2 at 95%, the highest anyone has scored against that board's 100% human panel, and sits fourth on LiveBench Mathematics at 96.81, behind Claude Opus 5.5 at 97.08, Claude Fable 5.1 at 97.01 and GPT-6.1 Sol at 96.83. It also reports 97.6% on FrontierMath Tier 4 v2, a number worth carrying with Epoch AI's disclosure that OpenAI funded the benchmark and holds exclusive access to part of it. The catch is price and reach: $10 / $50 per 1M tokens against GPT-6.1 Sol's $2 / $10, and it needs a paid ChatGPT plan. GPT-6.1 Sol, released September 29, is the value alternative and very nearly a co-leader: 92.63 on LiveBench Reasoning, 0.02 behind Astra, and third on LiveBench Mathematics, at a fifth of Astra's token price. ARC Prize has not rated it yet, which is why the crown stays with Astra. GPT-5.6 Sol, the old value pick, is now fifth on LiveBench Reasoning and third on ARC-AGI-2 at 92.5%. Qwen 3.7 Max is the budget pick for competition-style problems at 97.1 on the February 2026 HMMT index and 44.5 on Apex, with 200 free model requests a day. Claude Opus 5 remains the alternative for long agentic reasoning chains, though Astra has taken ARC-AGI-3 from it as well.

Model	Best For	Key Benchmark	Weakness	Price
GPT-6 Astra	Hardest math, science and reasoning	#1 LiveBench Reasoning (92.65), #1 ARC-AGI-2 (95%), #1 ARC-AGI-3	Needs a paid ChatGPT plan; 5x GPT-6.1 Sol's price	$10 / $50
Claude Opus 5	Long agentic reasoning chains	1673 on AA-Briefcase v1.1 and 1708 on GDPval-AA v2.1, fourth among distinct models on both; 30.2% ARC-AGI-3	Astra now leads ARC-AGI-3; 50% hallucination rate; now a legacy model	$5 / $25
GPT-5.6 Sol	Value STEM flagship	#3 ARC-AGI-2 (92.5%); #5 LiveBench Reasoning (91.65), Mathematics 96.20	FrontierMath still unpublished; scheming flagged by METR	$100/mo ChatGPT Pro; API $4 / $20
GPT-6.1 Sol	Value pick, near-Astra reasoning	#2 LiveBench Reasoning (92.63), #3 LiveBench Mathematics (96.83); Intelligence Index 51.8 at $0.72 an index task	Not yet rated by ARC Prize; in ChatGPT Work and Codex, not in Chat	$2 / $10
Qwen 3.7 Max	Competition math on a budget	97.1 HMMT 2026 Feb, 44.5 Apex, 200 free requests/day	API-only	$1.25 / $3.75 promo; $2.50 / $7.50 list
Claude Fable 5	Math inside a coding workflow	#1 on Arena's math subcategory (1543), 96.0 LiveBench Mathematics	Priciest option here	$10 / $50
GLM 5.3 Flash	Open-weight problem solving	Intelligence Index 41.8 under MIT, almost eight points above GLM-5.2 at less than half the size; 320B/18B active, 1M context	Needs a multi-GPU server, not a workstation; GLM 5.3 scores higher on Z.ai's own figures and is downloadable since August 28, but at more than twice the size and under a bespoke licence	Open weights (MIT)

Runner-up and alternatives: GPT-6.1 Sol is the runner-up and the value pick at $2 / $10, Claude Opus 5 is the natural pick for long-chain agentic reasoning, Qwen 3.7 Max is the budget pick, and GLM 5.3 Flash is the open-weight pick. Our dedicated guide to the best AI for math covers the task-by-task split and the free options.

What changed this month

The problem-solving crown moved to GPT-6 Astra, which took the boards this category is decided on within days of its September 3 launch. It leads LiveBench Reasoning at 92.65 against GPT-5.6 Sol's 91.65, takes ARC-AGI-2 at 95% against Sol's 92.5%, and displaced Claude Opus 5 on ARC-AGI-3, where its 62.7% on the standard harness beats Opus 5's 30.2%. Read the ARC-AGI-3 figures carefully: that board measures human-level action efficiency on unseen environments rather than problems solved, and the widely quoted 98.6% comes from a provider-adapter harness rather than the standard one. Sol drops to third on both LiveBench boards but stays the value alternative at $4 / $20 against Astra's $10 / $50, and LiveBench Mathematics has since changed hands twice, to Claude Fable 5.1 at 97.01 and then to Claude Opus 5.5 at 97.08 on September 22. Two things moved after that. ARC Prize scored Claude Opus 5.5 at 93.3% on ARC-AGI-2, which drops GPT-5.6 Sol to third on that board, and GPT-6 Sol arrived on September 22 at $2 / $10, half what the 5.6 tier costs, with a higher Intelligence Index but no LiveBench or ARC Prize score of its own yet. Claude Sonnet 5.5, released September 28, does not move this crown either: it scores 91.63 on LiveBench Reasoning, fifth and a hair behind GPT-5.6 Sol, and ARC Prize has not rated it yet. GPT-6.1 Sol arrived on September 29 and came closest yet: 92.63 on LiveBench Reasoning, 0.02 behind Astra, and 96.83 on Mathematics, which drops Astra to fourth there. It takes the podium's third place and the value pick from GPT-5.6 Sol, which is now fifth on LiveBench Reasoning rather than third. The crown stays with Astra because it still leads Reasoning and ARC-AGI-2, and ARC Prize has not rated GPT-6.1 Sol.

Category pick, September 2026

Best AI Agent

Gemini Spark

Cloud-resident

→

Claude Cowork

Desktop-resident

→

OpenAI dots

Always-on, runs on GPT-6 Astra

The best AI agent is Gemini Spark if you want the agent in the cloud and Claude Cowork if you want it on your desktop. These are joint picks rather than a first and a second: Spark runs in a Google Cloud VM, so it keeps working while your laptop is shut, wired into Gmail, Docs and, since early September, Google Photos; Cowork runs on your Mac or Windows machine and drives your local apps and your screen. Spark is no longer the only agent that keeps working in the cloud. SpaceXAI's Grok Bot (August 11) gives each account a persistent cloud computer that signs into your apps, from a $20/month Cursor Pro plan; Meta's Muse (September 8) runs errands from a cloud VM of its own and is free up to 100 million tokens a week; and OpenAI dots (September 29) run on GPT-6 Astra with their own cloud computer and browser, from the $100/month Pro plan, which does not include them in the EEA, UK or Switzerland. No independent board rates these products against each other, so the choice is about where the work has to happen rather than which one scores higher. Price is no longer a tiebreaker either: Spark now reaches the $19.99/month Google AI Pro tier in the US and more than 160 other countries, and Cowork is included at no extra charge on every paid Claude plan from $20/month. ChatGPT Codex Mobile (May 14) is the alternative for coding-agent work. Read the full Gemini Spark vs Claude Cowork comparison.

Agent	Best For	Where It Runs	Strength	Price
Gemini Spark	24/7 cloud tasks, Workspace workflows	Google Cloud VM (always-on)	Always-on, deep Workspace integration	From $19.99/mo Google AI Pro
Claude Cowork	Desktop, app-driving, design + code	Your Mac/Windows desktop	Drives local apps, sees your screen	$20/mo Claude Pro
ChatGPT Codex Mobile	Coding agent on phone	OpenAI cloud + iOS/Android	Approve diffs and redirect work from phone	Included in ChatGPT plans
Grok Bot	Browser jobs across logged-in apps	One persistent SpaceXAI cloud computer shared by all your Bots	Signs into apps with your own logins and keeps going with your laptop closed	From $20/mo Cursor Pro, or a linked SuperGrok plan
OpenAI dots	Ongoing responsibilities, not one-off tasks	OpenAI cloud computer and browser	Runs on GPT-6 Astra; keeps working with your computer off; reachable in ChatGPT, Slack and by voice	From $100/mo ChatGPT Pro (not EEA/UK/CH); Business Premium
Meta Muse	Free personal errands	A Muse Secure VM in Meta's cloud	Free up to 100M tokens a week; a separate Sentinel agent approves what it sends out	Free; $20/mo Power, $100/mo Maximum

Runner-up and alternatives: Gemini Spark and Claude Cowork are joint picks split by where the agent runs, not a first and a second. OpenAI dots are the pick if you already pay for ChatGPT Pro and want an agent on OpenAI's flagship, Grok Bot is the pick for browser jobs across your logged-in apps, from a $20/month Cursor Pro plan, Meta's Muse is the one to try for free, and ChatGPT Codex Mobile is the pick for coding agents. Google AI Ultra at $99.99 or $199.99 a month now buys Spark higher usage limits rather than access, which starts on the $19.99 Pro tier.

What changed this month

Three always-on cloud agents now compete with Gemini Spark, SpaceXAI's Grok Bot (August 11), Meta's Muse (September 8) and OpenAI dots (September 29), and this block now lists all three. No independent board rates them against each other, so the crowns do not move: Gemini Spark (cloud) and Claude Cowork (desktop) remain joint picks split by where the agent runs. Spark did expand, reaching the $19.99/month Google AI Pro tier and gaining Google Photos, where it can search, edit, curate and run scheduled workflows. The model layer underneath moved a long way. Claude Opus 5.5 (September 22) leads both agentic boards Artificial Analysis publishes, AA-Briefcase v1.1 at 1822 and GDPval-AA v2.1 at 1846, and Claude Sonnet 5.5 (September 28) is second on both at 1811 and 1844, ahead of Claude Fable 5.1 at 1678 and 1735. Sonnet 5.5 also leads Artificial Analysis's AutomationBench at 71.3% against Opus 5.5's 69.5%, at half its token price. Opus 5.5 is the model most teams should build on, and Sonnet 5.5 is the cheaper one to try first on well-scoped agent work. Arena's Agent Arena has now rated Opus 5.5 and puts it second behind Claude Fable 5.1, which still leads on confirmed task success at 17.4% against Opus 5.5's 16.3% and GPT-6 Astra's 12.7%; it has not rated Sonnet 5.5 yet. SpaceXAI's Grok 4.7 (September 21) is fifth among distinct models on both of Artificial Analysis's agentic boards, at 1695 and 1657, above GPT-6 Astra on both, though it reaches the API, Cursor and Grok Build rather than a consumer agent product. Meta's Muse Spark 1.3 sits seventh among distinct models on GDPval-AA v2.1 at 1672, and Scale SEAL has still rated only the older 1.1, which leads its MCP Atlas tool-use board at 88.1. GLM 5.3 Flash (August 26, MIT) is still the open-weight agent model we would host, at AutomationBench 48.8 on Z.ai's own chart where Claude Opus 4.8 scores 41.0. On the Agent Arena, Tencent's Hy4 Preview is now the highest-placed open model at twelfth, just ahead of Kimi K3, with DeepSeek V4.1-Flash fifteenth and GLM-5.2 sixteenth. GPT-6.1 Sol (September 29) scores 1575 on GDPval-AA v2.1, above GPT-6 Astra's 1542 at a fifth of the token price.

Fello AI running on Mac, iPhone and iPad

All the top AI models, one app

The leading models like ChatGPT, Claude and Gemini, together on Mac, iPhone and iPad.

Free to start, 4.7★ across 27,000+ reviews.

Download Fello AI

Use case guide, September 2026

Best AI for Students

GPT-5.6 Luna

Essays & research

→

Gemini 3.6 Flash

STEM + PDFs

→

Claude Sonnet 5.5

Writing & editing

The best AI for students is GPT-5.6 Luna inside ChatGPT for general coursework and Gemini 3.6 Flash inside the Gemini app for STEM and multimodal study, with Qwen 3.7 Max as the API alternative for harder problem sets (200 free requests a day) and Claude Sonnet 5.5 as the free alternative for essay editing. Most students don't need to pay, and the free tier improved on August 6: GPT-5.6 Luna became the default for ChatGPT Free and Go, with unlimited text chats and a Think button for harder questions, though file uploads and image tools stay capped. On September 22 OpenAI added GPT-6 Luna for Free and Go users in the ChatGPT desktop app, a current-generation model on the free tier from launch day, though it scores the same 37.3 as GPT-5.6 Luna on Artificial Analysis's index. Gemini 3.6 Flash is still what the free Gemini app serves, Claude Sonnet 5.5, released September 28, is free on claude.ai, and DeepSeek V4 is free on DeepSeek's chat site. Luna is the cheapest member of the GPT-5.6 family rather than the strongest, so reach for the Think button or switch to GPT-5.5 when an essay needs more care. For step-by-step working on the hardest math, GPT-5.6 Sol is OpenAI's paid flagship and GPT-6 Astra now reports 97.6% on FrontierMath Tier 4 v2 against GPT-5.5 Pro's verified 39.6%, though Astra is not on ChatGPT's free tier; Qwen 3.7 Max is the value alternative at 97.1 HMMT 2026 February with API pricing at $1.25 / $3.75 on its current 50% promo ($2.50 / $7.50 list).

Task	Best Model	Why	Free?	Alternative
Essays & coursework	GPT-5.6 Luna	Free default in ChatGPT since August 6; unlimited text chats, plus GPT-6 Luna in the desktop app since September 22	Yes	Claude Sonnet 5.5 (free Claude)
STEM problem-solving	GPT-5.6 Sol / Qwen 3.7 Max	Paid STEM flagship; Astra reports 97.6% FrontierMath Tier 4 v2 / 97.1 HMMT 2026 Feb	Pro paid / Qwen API paid	Gemini 3.6 Flash (free)
Research & accuracy	Gemini 3.1 Pro	98% ARC-AGI-1 at $0.52/task, native Google Search grounding	Yes (Gemini app)	Claude Opus 5
Writing editing	Claude Sonnet 5.5	Free on claude.ai since September 28 and second only to Opus 5.5 on the Intelligence Index; Claude Fable 5.1 is the quality leader	Yes (Claude free)	Claude Fable 5.1
Multimodal study (PDFs, slides, images)	Gemini 3.6 Flash	1M context, free in Gemini app	Yes	NotebookLM (Google)

Runner-up and alternatives: Claude Sonnet 5.5 (free) is the runner-up for essay writing and editing. Gemini 3.6 Flash (free) is the runner-up for multimodal study and PDF ingestion. DeepSeek V4 is the runner-up for problem-solving on a strict zero-cost budget.

What changed this month

The free tier is where this guide moved. On August 6 OpenAI made GPT-5.6 Luna the default for ChatGPT Free and Go and gave both unlimited text chats plus a Think button for harder questions, so the free pick is now Luna rather than GPT-5.5, which stays selectable when an essay needs more care. File uploads, images and other tools are still capped. Nothing moved on the Google side: Gemini 3.8 Flash shipped on September 2 but reaches free users only through AI Studio and the API, so the free Gemini app still serves 3.6 Flash and it keeps the multimodal row. GPT-6 Astra (September 3) is the one to watch for hard problem sets, reporting 97.6% on FrontierMath Tier 4 v2 against GPT-5.5 Pro's verified 39.6%, but it needs a paid ChatGPT plan and no free tier includes it. The free tier moved again on September 22: Free and Go users can now reach GPT-6 Luna in the ChatGPT desktop app. Read it as cheaper rather than smarter, because Artificial Analysis scores it 37.3, level with the 5.6 build, and ARC Prize puts the two within two tenths of a point of each other on ARC-AGI-2. Claude's free model moved on September 28: Sonnet 5.5 is available on the Free plan at claude.ai and scores 56.0 on Artificial Analysis's index against Sonnet 5's 38.2, so it replaces Sonnet 5 as the writing and editing pick. It runs at Medium effort in the apps, and free-plan usage limits still apply.

Use case guide, September 2026

Best AI for Work & Professionals

GPT-5.6

Daily knowledge work

→

Claude Opus 5.5

Coding & writing

→

Gemini 3.1 Pro

Research & briefings

The best AI for professional work is GPT-5.6 (ChatGPT's default since July 9) for daily knowledge work, Claude Opus 5.5 for coding and high-stakes writing, and Gemini Spark for 24/7 agentic workflows. Most professionals get the most out of running two paid subscriptions (ChatGPT Plus at $20/month plus Claude Pro at $20/month, total $40/month), or consolidating with Fello AI at $9.99/month for all five top models in one Mac/iOS app. For agentic work that runs while you sleep, Gemini Spark is the always-on cloud agent that lives inside Google Workspace, and since July 30 it reaches the $19.99/month Google AI Pro tier as well as Google AI Ultra; OpenAI dots, Grok Bot and Meta's Muse are the alternatives, covered in the agents section above.

Use Case	Best Model	Key Stat	Price	Alternative
Daily knowledge work	GPT-5.6	ChatGPT's default since July 9; the assistant most people can open	$20/mo ChatGPT Plus	Claude Opus 5.5
ChatGPT Work & Codex	GPT-6.1 Sol	In ChatGPT Work and Codex for Plus, Pro, Business, Enterprise and Edu since September 29, at $2 / $10 in the API	$20/mo ChatGPT Plus	GPT-6 Luna for volume work
Coding (proprietary)	Claude Opus 5.5	#1 Arena WebDev (1820) and #1 Intelligence Index (57.6), at $4 / $20	$20/mo Claude Pro	Claude Sonnet 5.5, which leads Terminal-Bench 4.0 at half the token price
Coding (cost-effective)	Claude Sonnet 5.5	#1 Terminal-Bench 4.0 (63.6%) and #1 LiveBench Coding (91.4); $1.08 an index task at high effort	$2 / $10	Qwen3.8-Max-0902; DeepSeek V4.1-Flash
Research & briefings	Gemini 3.1 Pro	98% ARC-AGI-1 at $0.52/task, Google grounding	Google AI Pro / Ultra	Claude Opus 5.5
Hard math, physics, finance modelling	GPT-6 Astra	#1 LiveBench Reasoning and ARC-AGI-2 (95%); reports 97.6% FrontierMath Tier 4 v2	$100/mo ChatGPT Pro	GPT-5.6 Sol; Qwen 3.7 Max
Always-on agent workflows	Gemini Spark	Always-on cloud agent inside Google Workspace	From $19.99/mo Google AI Pro	Claude Cowork
Live news, X-context creative	Grok 4.7	Intelligence Index 46.3 + native X grounding; the Grok app still serves 4.6	$30/mo SuperGrok	Gemini 3.1 Pro
All-in-one consolidation	Fello AI	ChatGPT + Claude + Gemini + Grok + DeepSeek	$9.99/mo	Pay each vendor separately

Runner-up and alternatives: for most professional teams, Claude Opus 5.5 is the runner-up to GPT-5.6 for daily work and the coding leader at $4 / $20, with Claude Sonnet 5.5 the cheaper coding pick that leads Terminal-Bench 4.0 and GPT-6 Astra the pick on Arena's image-to-WebDev board. Gemini 3.1 Pro is the runner-up for research-heavy roles, and Gemini Spark is the pick if you can put a cloud agent to work on long tasks inside Google Workspace, with OpenAI dots the alternative for ChatGPT Pro subscribers.

What changed this month

Three launches landed in the first three days of September and a fourth on September 22, and the last one does change the price argument this block rests on. Claude Opus 5.5 tops Artificial Analysis's Intelligence Index at 57.6 on v4.3.2 and both of its agentic boards, and it does that at $4 / $20 against Claude Opus 5's $5 / $25. The reasoning this block used to carry, that teams should start with Opus 5 because it costs half what the better models cost, no longer applies: Opus 5.5 is both better and cheaper, so it takes the proprietary coding row and is the model we would build daily work on. GPT-5.6 keeps the daily-assistant row all the same, because that row is about reach rather than score and neither Opus 5.5 nor GPT-6 Astra is a default for hundreds of millions of people. Claude Sonnet 5.5 (September 28) takes the cost-effective coding row from Alibaba's Qwen3.8-Max-0902: it leads Terminal-Bench 4.0 and LiveBench Coding at $2 / $10, and at high effort an index task costs $1.08. Keep it below max effort, where the same task costs $7.60, more than on Opus 5.5. GPT-6 Sol and GPT-6 Luna landed on September 22 in exactly the surface this block is about: ChatGPT Work and Codex, for Plus, Pro, Business, Enterprise and Edu, at $2 / $10 and $0.10 / $0.50 in the API, and not in the ordinary chat window. GPT-6.1 Sol replaced GPT-6 Sol in the ChatGPT Work and Codex row on September 29, at the same $2 / $10 and with a higher score, 51.8 against 47.5, and OpenAI dots gave ChatGPT Pro its own always-on agent the same day.

Pricing Comparison

AI Model Pricing in September 2026

From $0 free tiers to $199.99/month Google AI Ultra. The most consequential price on this table is Claude Opus 5.5 at $4 / $20, because on September 22 Anthropic shipped the highest-scoring model any independent evaluator has measured at a lower list price than the model it replaces: Claude Opus 5 is $5 / $25, and cache reads fall 60% to $0.20, with Anthropic's own figure putting a typical workload 40% cheaper than on Opus 5. That inverts the shape this section has had all year, in which the best model was also the most expensive. Claude Fable 5.1 and GPT-6 Astra both list at $10 / $50 and both now score below it. Six days later Anthropic added Claude Sonnet 5.5 at $2 / $10, second only to Opus 5.5 on the index, but check the effort setting before counting the saving: at max effort an index task costs $7.60, more than Opus 5.5's $5.98, while at high effort it costs $1.08. Meta's Muse Spark 1.3 lists at $1.25 / $4.25, unchanged across three capability bumps since July, with a Contributor tier at $0.10 / $0.20 for anyone willing to let Meta train on their prompts, and Grok 4.6 and Grok 4.7 both list at $2 / $6, with the caveat that a prompt of 200,000 tokens or more re-bills the entire request at $4 / $12. Grok 4.7 is the clearest case on this page of a price that did not move while the cost did: the tokens cost the same, and an Intelligence Index task costs $2.73 against Grok 4.6's $1.86, because 4.7 emits roughly twice the output getting there. The cheap end has moved twice. DeepSeek raised both V4 models roughly fourfold on August 16, which cost V4-Flash the price-performance pick, then reversed most of it on September 10: V4-Flash was retired and V4.1-Flash took its place at $0.15 / $0.60 off-peak and $0.30 / $1.20 at peak. Among models you can buy by the token the pick stays with GLM 5.3 Flash, now at its plain $0.15 / $0.50 list since the launch promotion expired on September 9, because Artificial Analysis measures it at $0.25 an Intelligence Index task for a score of 41.8 against DeepSeek's $0.27 for 39.5. Off-peak the two are level on input and GLM is cheaper on output; at peak GLM wins both. One open model beats them both on cost per task and it is four days old: Xiaomi's MiMo-V2.6-Pro runs an index task for $0.13 at a score of 46.3, under plain MIT, but you have to host 1.02 trillion parameters to get it. At the frontier the value pick is Meta's Muse Spark 1.3, at $1.60 an index task for 48.1. Among closed models GPT-6 Luna took that title on September 22, at $0.10 / $0.50 by the token and $0.07 an index task, the cheapest per task of anything on this page; GPT-6 Sol fell the same way, from $4 / $20 to $2 / $10, landing exactly on Claude Sonnet 5.5's rate, and OpenAI says both cuts are permanent rather than promotional. GPT-6.1 Sol arrived on September 29 at that same $2 / $10 with cached input at $0.10, and at $0.72 an index task it is the cheapest route to a top-five Intelligence Index score. For a deeper breakdown see our full AI Pricing Comparison Guide.

Model	Input (per 1M)	Output (per 1M)	Context	Free Access?
GPT-6 Astra	$10.00	$50.00	1,050,000 (922,000 max input)	In Chat on Pro, Business and Enterprise since September 4; Plus gets it in ChatGPT Work and Codex rather than Chat; no free tier; live in the API
GPT-6 Sol	$2.00	$10.00	1,050,000 (922,000 max input)	ChatGPT Work & Codex on Plus, Pro, Business, Enterprise and Edu since September 22; API; not in Chat
GPT-6.1 Sol	$2.00	$10.00	1,050,000 (922,000 max input)	ChatGPT Work & Codex on Plus, Pro, Business, Enterprise and Edu since September 29; API; not in Chat; cache reads $0.10
GPT-6 Luna	$0.10	$0.50	1,050,000 (922,000 max input)	Free and Go in the ChatGPT desktop app; ChatGPT Work & Codex on paid plans; API; not in Chat
GPT-5.5	$5.00	$30.00	1M (400K in Codex)	ChatGPT Free; API paid
GPT-5.5 Pro	$30.00	$180.00	1M	ChatGPT Pro from $100/mo ($200 and $500 higher-usage tiers)
GPT-5.6 Sol	$4.00	$20.00	Not published	Live in ChatGPT, Codex & API (July 9); promotional rate through at least November 21, 2026
GPT-5.6 Terra	$2.00	$12.00	Not published	Live in ChatGPT, Codex & API (July 9)
GPT-5.6 Luna	$0.20	$1.20	Not published	Live in ChatGPT, Codex & API (July 9)
Claude Opus 5.5	$4.00	$20.00	1M	Claude Pro/Max/Team; API paid; cache reads $0.20
Claude Opus 5	$5.00	$25.00	1M	Legacy model at Anthropic; Pro/Max/API
Claude Opus 4.8	$5.00	$25.00	1M	Legacy model at Anthropic; Pro/Max/API
Claude Fable 5.1	$10.00	$50.00	1M	Cache reads $0.25; in Max/Team Premium (~50% of usage limits); Pro/Team Standard via credits. Mythos 5.1 bills identically, by invitation only
Claude Fable 5	$10.00	$50.00	1M	Permanent in Max/Team Premium (~50% of usage limits); Pro/Team Standard via credits
Claude Sonnet 5.5	$2.00	$10.00	1M	Free on claude.ai; Pro/Max/Team; API paid; cache reads $0.20
Claude Sonnet 5	$2.00	$10.00	1M	Legacy model since September 28; the fallback for flagged cyber requests; API paid
Claude Sonnet 4.6	$3.00	$15.00	1M	API paid (superseded by Sonnet 5)
Gemini 3.1 Pro	$2.00 (≤200K) / $4.00 (>200K)	$12.00 (≤200K) / $18.00 (>200K)	1M	Limited Gemini app; API paid
Gemini 3.8 Flash	$0.75 intro / $1.50 from Jan 1, 2027	$3.75 intro / $7.50 from Jan 1, 2027	1M	AI Studio, Antigravity, Gemini Enterprise + paid API; in the Gemini app reported for AI Pro/Ultra
Gemini 3.7 Flash	$0.75 intro / $1.50 from Jan 1, 2027	$3.75 intro / $7.50 from Jan 1, 2027	1M	AI Studio, Android Studio, Antigravity + paid API; in the Gemini app via Spark only (AI Pro/Ultra, excludes EEA/UK/CH/Nigeria)
Gemini 3.6 Flash	$0.75 intro / $1.50 from Jan 1, 2027	$3.75 intro / $7.50 from Jan 1, 2027	1M	Free Gemini app default; AI Studio; free API tier + paid API
Gemini 3.5 Flash-Lite	$0.30	$2.50	1M	AI Studio; free API tier + paid API
Qwen3.8-Max-0902	$2.00 ($0.17 explicit cache hit, $0.25 implicit)	$6.00	1M (128K output)	QwenCloud API only; no consumer chat front-end, no open weights
Qwen 3.7 Max	$1.25 promo / $2.50 list	$3.75 promo / $7.50 list	1M	200 free requests/day; API paid beyond that
MiniMax M3	$0.30 (50% off $0.60)	$1.20 (≤512K)	1M	Open weights; hosting costs apply
LongCat-2.0	Provider-dependent	Provider-dependent	1M	Open weights (MIT); hosting costs apply
NVIDIA Nemotron 3 Ultra	Provider-dependent	Provider-dependent	1M	Open weights (OpenMDW); hosting costs apply
Qwen 3.5 (open-weight)	Self-host / Together	Self-host / Together	1M	Open weights; hosting costs apply
Nex-N2-Pro	Self-host / providers	Self-host / providers	1M	Open weights (Apache 2.0); hosting costs apply
Rio 3.5 Open 397B	Self-host / providers	Self-host / providers	1M	Open weights (MIT); hosting costs apply
Grok 4.3	$1.25	$2.50	1M	Free consumer plan; API paid
Grok 4.6	$2.00 (<200K) / $4.00 (≥200K)	$6.00 (<200K) / $12.00 (≥200K)	500K	SpaceXAI API, Cursor, Grok Build, OpenRouter, Vercel, Cloudflare
Grok 4.7	$2.00 (<200K) / $4.00 (≥200K)	$6.00 (<200K) / $12.00 (≥200K)	500K	SpaceXAI API, Cursor, Grok Build, harnesses and cloud routers; not in the Grok app
Muse Spark 1.3	$1.25 ($0.10 Contributor tier)	$4.25 ($0.20 Contributor tier)	1M	Paid API; Muse Code, Meta Model API and OpenRouter; Contributor tier trades training rights for a ~92% discount
Kimi K3	$3.00 ($0.30 cache-hit)	$15.00	1M	Free basic tier in the Kimi app; open weights on Hugging Face (Kimi K3 License)
Gemini Omni Flash (video)	$1.50	$17.50 (video output)	10-second clips	Gemini app / Flow; AI Studio + API
DeepSeek V4-Pro	$0.66 off-peak / $1.32 peak ($0.022 cache-hit off-peak)	$1.98 off-peak / $3.96 peak	1M	DeepSeek Chat free; API paid, peak/off-peak tiers since August 16
DeepSeek V4.1-Flash	$0.15 off-peak / $0.30 peak ($0.003 cache-hit off-peak)	$0.60 off-peak / $1.20 peak	1M	DeepSeek Chat free; API paid, peak/off-peak tiers; replaced V4-Flash on September 10
Kimi K2.7 Code	Provider-dependent	Provider-dependent	256K	Open weights; hosting costs apply
GLM-5.2	Provider-dependent	Provider-dependent	1M	Open weights; hosting costs apply
GLM 5.3	$1.40 ($0.26 cache-hit)	$4.40	1M	Z.ai API, GLM Coding Plan and ZCode; weights on Hugging Face since August 28 under the bespoke GLM 5.3 License
GLM 5.3 Flash	$0.15 ($0.03 cache-hit); $0.075 promo	$0.50; $0.25 promo	1M	Z.ai API, GLM Coding Plan, OpenRouter; MIT weights on Hugging Face; promo ends September 9
Tencent Hy4 Preview	$0.834 ($0.042 cache-hit)	$2.501	1M+	Open weights (Apache 2.0); Tencent Cloud TokenHub, OpenRouter, WorkBuddy, CodeBuddy
Qwen3.8-Flash-Next	Provider-dependent	Provider-dependent	262K (to 1M)	Open weights (Qwen Community License 1.0); hosting costs apply
MiMo-V2.6-Pro	$0.435	$0.87	1M	Open weights (MIT); hosting costs apply
ERNIE 5.1	China-region pricing	China-region pricing	256K	Baidu free tier
Gemini Spark (agent)	Not API-priced	Not API-priced	1M (Gemini base)	Google AI Pro $19.99, AI Ultra $99.99 or $199.99/mo
Fello AI (aggregator)	Routed via app	Routed via app	Model-dependent	$9.99/mo, free tier available

The GPT-5.5 and GPT-5.5 Pro rates above are short-context prices; OpenAI no longer publishes the specific long-context figures. The GPT-5.6 tiers are billed at 2x input and 1.5x output once a prompt passes 272K input tokens, which puts long-context Terra at $4 / $18 and Luna at $0.40 / $1.80. The GPT-6 tiers carry the same surcharge past 272K input tokens, applied to the whole request, and their cached input reads are 90% off: $1.00 for Astra, $0.20 for GPT-6 Sol and $0.01 for Luna, with GPT-6.1 Sol at 95% off, $0.10. Grok 4.6 works the same way but bites harder: at 200,000 prompt tokens and above, SpaceXAI re-bills the entire request at the higher rate rather than only the overage, so crossing the line by a thousand tokens doubles the cost of the whole call. DeepSeek is the other rate card that needs reading twice: since August 16 both V4 models bill at peak rates from 01:00 to 04:00 and from 06:00 to 10:00 UTC on weekdays, and at exactly half that in every other hour, so the same job can cost twice as much depending on when you run it. If you want access to multiple AI models without managing separate subscriptions, Fello AI provides GPT, Claude, Gemini, Grok, Perplexity, and more in a single app for Mac, iPhone, and iPad from $9.99/month.

Open-Weight Models

Best Open-Weight Models in September 2026

The best open-weight model in September 2026 is MiMo-V2.6-Pro, which Xiaomi published on September 21 under plain MIT: 1.02 trillion parameters with 42 billion active, a 1M-token context, and the weights on Hugging Face on the day. Artificial Analysis measures it at 46.3 on Intelligence Index v4.3.2, which is the highest open score it has ever published, above GLM-5.3 at 44.8 and Kimi K3 at 43.6, and it does that at $0.13 an Intelligence Index task, cheaper than any other open model on this page. It also takes 34.8% on Terminal-Bench 4.0 and 1678 on GDPval-AA v2.1, which puts it above Kimi K3 on both. Two honest limits: Arena's human voters are cool on it, 21st on WebDev and 23rd on text, and it has barely a week of independent track record. Kimi K3 is still the highest-placed open model on Arena, tenth on WebDev at 1659, though on the Agent Arena it now sits thirteenth, one place behind Tencent's Hy4 Preview, and GLM-5.3 is still the strongest open model on Artificial Analysis's Terminal-Bench 4.0 at 41.9% against MiMo's 34.8%, on a bespoke Z.ai licence rather than MIT. The practical recommendation for teams running their own weights is still GLM 5.3 Flash (Z.ai, MIT), released on August 26 and scoring 41.8 on v4.3.2, 320B total with 18B active, because a 1.02-trillion-parameter model is not a workstation proposition and the K3 download is 96 safetensors shards and about 1.56 TB. The cheap open tier changed again on September 10: DeepSeek retired V4-Flash 0731 and replaced it with DeepSeek-V4.1-Flash, 552B with 8B active at prefill and 16B at decode, natively multimodal, MIT on the day, scoring 39.5 on v4.3.2 against 34.3 for the build it replaces, and priced back down to $0.15 / $0.60 off-peak. It also has the highest confirmed task-success rate of any open model on the Agent Arena, 12.1%, though it ranks fifteenth there overall. Of the three releases that closed August, GLM 5.3 published its weights on August 28 under a bespoke licence with a clause requiring any model-as-a-service operator above $10 billion in revenue to pass a Z.ai security review first; Alibaba's Qwen3.8-Flash-Next, a 125B mixture-of-experts with only 6B active that previews the Qwen4 architecture, scores 39.8 and sits twelfth on Arena's WebDev board; and Tencent's Hy4 Preview, 770B total and 49B active under Apache 2.0, has no Artificial Analysis rating but sits fourteenth on Arena WebDev at 1633. Note also that GLM 5.3 Flash is not a trimmed GLM 5.3 but a separate model on a newly trained base, which is why it could ship under plain MIT while the flagship could not.

Model	Best For	Key Benchmark	Context / License	Where To Run
MiMo-V2.6-Pro	Highest open Intelligence Index ever measured	II 46.3 (v4.3.2), above GLM-5.3 and Kimi K3, at $0.13 an index task; 34.8% Terminal-Bench 4.0; GDPval-AA v2.1 1678; 1.02T/42B active	1M / MIT	Hugging Face (XiaomiMiMo), providers, self-host
Kimi K3	Highest-placed open model on Arena	II 43.6 (v4.3.2); #10 Arena WebDev, the best open placement there; #13 Agent Arena; 2.8T/104B active	1M / Kimi K3 License	Hugging Face (96 shards, 1.56 TB), Moonshot API, providers
GLM 5.3 Flash	Best open model most teams can actually run	II 41.8 (v4.3.2), on the intelligence-versus-cost Pareto frontier at about $0.25 an index task; 320B/18B active, natively multimodal	1M / MIT	Hugging Face, Z.ai API ($0.15/$0.50), OpenRouter
GLM-5.2	Fallback with an independent track record	II 33.7 (v4.3.2); #23 Arena WebDev (1,604.3), #16 Agent Arena; 744B/40B active	1M / MIT	Z.ai, Hugging Face, OpenRouter
GLM 5.3	Open since August 28, but on a bespoke licence	II 44.8 (v4.3.2), the second-highest open score we found; same 744B base as GLM-5.2, post-trained only, published checkpoint counted at 753B; CyberGym 84.5% and Terminal-Bench 3.0 28.3 (vendor figures)	1M / GLM 5.3 License (MaaS above $10B revenue needs a Z.ai security review)	Hugging Face, Z.ai API ($1.40/$4.40), GLM Coding Plan, ZCode
DeepSeek V4.1-Flash	Best open value if you host it yourself	II 39.5 (v4.3.2) against 34.3 for the V4-Flash 0731 build it replaces; 552B, 8B active at prefill and 16B at decode	1M / MIT	DeepSeek API ($0.15/$0.60 off-peak, $0.30/$1.20 peak since September 10), local
Tencent Hy4 Preview	Largest permissively licensed model yet	770B/49B active; 2.99/4.00 on Tencent's own 203-task expert panel vs Kimi K3 2.94 and GLM 5.3 2.92 (vendor); no Artificial Analysis rating; #14 Arena WebDev (1633), #12 Agent Arena	1M+ / Apache 2.0	Hugging Face (BF16 and FP8), Tencent Cloud TokenHub, OpenRouter
Qwen3.8-Flash-Next	Preview of the Qwen4 architecture	125B total plus a 51B N-gram embedding, 6B active; 91.7 GPQA Diamond and 62.5 SWE-Bench Pro (vendor); II 39.8 (v4.3.2); #12 Arena WebDev (1637)	262K to 1M / Qwen Community License 1.0	Hugging Face, providers, self-host
LongCat-2.0	Frontier open coder trained on Chinese chips	II 33 (v4.1); 59.5% SWE-Bench Pro (vendor), 1.6T/~48B active	1M / MIT	Hugging Face, GitHub, OpenRouter
MiniMax M3	Cheap frontier-class multimodal	II 44 (v4.1), 59% SWE-Bench Pro, multimodal	1M / license TBD	Hugging Face, API $0.30/1M (50% off)
Nex-N2-Pro	Strongest open coding score	II 41 (v4.1); 80.8 SWE-Bench Verified, 397B/17B active	Qwen-based / Apache 2.0	Hugging Face, providers, self-host
Kimi K2.7 Code	Strongest commercially-licensed open coder	+21.8% on Kimi Code Bench v2 vs K2.6 (vendor); 1T/32B active	256K / Modified MIT	Hugging Face, DeepInfra, providers
DeepSeek V4-Pro	Agentic real-world work	II 44 (v4.1), 1.6T/49B active	1M / MIT	DeepSeek API ($0.66/$1.98 off-peak, $1.32/$3.96 peak since August 16), local
Hy3	Newest permissive-licence entrant	II 41 (v4.1); #49 on Arena WebDev (1,508.6)	Apache 2.0	Hugging Face, providers, self-host
Inkling	Thinking Machines' first open model	II 41 (v4.1), agentic 32.3, released July 15	Open weights	Hugging Face, providers, self-host
Inkling Small	Same family at a quarter the size	II 40 (v4.1), agentic 30.8; 276B/12B active, text, image and audio in	Apache 2.0	Hugging Face (BF16 and NVFP4), providers, self-host
NVIDIA Nemotron 3 Ultra	NVIDIA-tuned, fully permissive license	II 38 (v4.1), 65-70.4 SWE-Bench Verified, 550B/55B active	1M / OpenMDW	OpenRouter, Hugging Face, AWS (8x B200 self-host)
Qwen 3.5 (397B / 17B active)	Multimodal, fast decode	88.4 GPQA, 91.3 AIME 2026, 83.6 LiveCodeBench v6	1M / open	Together, OpenRouter, local
Qwen3.6-35B-A3B	Efficient open agentic coder (3B active)	86.0 GPQA Diamond, 92.7 AIME 2026, 35B/3B active	262K (to 1M YaRN) / Apache 2.0	Hugging Face, OpenRouter, local
Qwen3.6-27B	Laptop-runnable dense coder	87.8 GPQA Diamond, dense 27B, multimodal	256K / Apache 2.0	Local Mac/PC, Hugging Face, OpenRouter
Rio 3.5 Open 397B	Qwen 3.5 fine-tune, multilingual reasoning	70.8 Terminal-Bench 2.1 (first-party), beats Qwen 3.7 Plus on 4/5	397B/17B active / MIT	Hugging Face, providers, self-host
Llama 4 Maverick	Meta-line flagship	17B active / 400B total params	Llama 4 license	Meta cloud, Hugging Face, local
NVIDIA Nemotron 3 Nano Omni	Edge / low-power	Multimodal, very small footprint	Compact / open	Local, NVIDIA tool

Licensing matters as much as raw score here, and August widened the gap between the two groups. Kimi K2.7 (Modified MIT), DeepSeek V4 (MIT), GLM 5.3 Flash (MIT), MiMo-V2.6-Pro (MIT), GLM-5.2 (MIT), LongCat-2.0 (MIT), Hy3 (Apache 2.0), Hy4 Preview (Apache 2.0), Nex-N2-Pro (Apache 2.0) and Nemotron 3 Ultra (OpenMDW) all clearly allow commercial use with no revenue test. The rest carry conditions worth reading before you build on them: MiniMax M3 and MiniMax H3 ship under their own community licences, H3 additionally requiring an application from the US, EU, UK and South Korea; Kimi K3 sits on a custom licence with a revenue threshold for anyone reselling it as a service; GLM 5.3 requires a Z.ai security review of any model-as-a-service operator above $10 billion in revenue; and the Qwen Community License 1.0 on Qwen3.8-Flash-Next requires a separate licence from Qwen to run a model-as-a-service or an AI work-assistant business, though internal use is exempt. No open model is first on any Arena board we track any more: Claude Opus 5 took the Agent board in July, the last one an open model led.

How We Evaluate

Benchmarks, Prices, and Hands-On Use

Every ranking on this page combines three inputs: public benchmarks from seven independent houses (Artificial Analysis, Arena formerly LMArena, Scale SEAL, LiveBench, EQ-Bench, ARC Prize and the official Terminal-Bench 4.0 board, covering the Intelligence Index, GPQA Diamond, ARC-AGI-1 through 3, Humanity's Last Exam, GDPval-AA, FrontierMath, HMMT, MCP Atlas, SWE Atlas and the Remote Labor Index), published API and subscription pricing from each vendor's official pricing page, and hands-on use by the FelloAI editorial team running real prompts across the same task on every model. We re-fetch official pricing and benchmark sources before every monthly update.

Benchmarks are weighted to the use case: SWE-bench and Terminal-Bench drive coding, GPQA Diamond and ARC-AGI-2 drive accuracy, GDPval-AA (Artificial Analysis's professional-deliverables benchmark) informs professional-task quality while writing style is judged primarily by hands-on testing, FrontierMath and HMMT drive problem-solving. We disclose when a benchmark is vendor-reported but not independently verified, and we strip any claim we cannot reproduce against a live source. We rank on measured scores: a category crown moves as soon as a model out-scores the holder on the boards that define that category, without waiting for the vote-based boards to catch up, and a model that is not yet widely available is flagged on its card rather than denied one. We re-check ranks as well as figures, because a score can stay correct while the board around it moves. When a model goes through a major upgrade between updates, we re-rank the category and add a "What changed this month" line at the bottom of the deep-dive.

Not sure which model to pick?

Fello AI brings the top AI models together in one lightweight native app.

Switch between them anytime, no separate subscriptions.

Download Fello AI

Frequently Asked Questions

Common Questions

What is the best AI model right now in September 2026?

It depends on the task, but Anthropic's two September releases now sit at the top of most boards. On overall benchmark score, Claude Opus 5.5 (September 22) is #1 on Artificial Analysis's Intelligence Index at 57.6 on v4.3.2 and Claude Sonnet 5.5 (September 28) is second at 56.0, both ahead of Claude Fable 5.1 at 53.4 and GPT-6 Astra at 52.7. Opus 5.5 also leads GDPval-AA v2.1 at 1846, AA-Briefcase v1.1 at 1822 and Humanity's Last Exam at 61.4%, at $4 / $20 per 1M tokens, and Arena now puts it first on its text, WebDev and creative-writing boards. For coding, Opus 5.5 holds the crown on Arena WebDev and LiveBench Agentic Coding, while Sonnet 5.5 leads Terminal-Bench 4.0 and LiveBench Coding at $2 / $10 and is the value pick. For daily chat, GPT-5.6 has been ChatGPT's default since July 9 and is the assistant most people can actually open. For writing, Claude Fable 5.1 has the best combined placing across the three prose boards. For accuracy, Claude Fable 5.1 still holds the highest factual accuracy Artificial Analysis has measured at 67% on AA-Omniscience, one point above Opus 5.5. For hard math and reasoning, GPT-6 Astra leads LiveBench Reasoning and ARC-AGI-2. For images, ChatGPT Images 2.5 leads all four boards we track, and for video Gemini Omni 1.1 Flash leads Arena text-to-video. For agents, Gemini Spark is the 24/7 cloud agent and Claude Cowork the desktop one. Among open weights, Xiaomi's MiMo-V2.6-Pro scores 46.3 under plain MIT at $0.13 an index task. GPT-6 Luna is the cheapest model per Intelligence Index task on this page at $0.07, and Claude Sonnet 5.5 is free on claude.ai. OpenAI's GPT-6.1 Sol (September 29) is fifth at 51.8 and the cheapest of the top five per index task, at $0.72.

Is GPT-6 out, and is it the best model now?

GPT-6 is out. OpenAI launched it as GPT-6 Astra on September 3, 2026, which settles the year-long question of whether Astra would ship as GPT-6 or as another GPT-5 point release. It is not the top model on the independent boards, and it slipped further on September 22. Artificial Analysis scores it 52.7 on Intelligence Index v4.3.2, 5.7 points clear of GPT-5.6 Sol but 0.7 behind Claude Fable 5.1 at 53.4, 3.3 behind Claude Sonnet 5.5 at 56.0 and 4.9 behind Claude Opus 5.5 at 57.6, while charging $10 / $50 per 1M tokens against Sol's $4 / $20 and Opus 5.5's $4 / $20. Coding was its strongest claim and it is now a split one: Astra took Artificial Analysis's Terminal-Bench 4.0 and held it until Opus 5.5 drew level at 59.6% and Claude Sonnet 5.5 went past at 63.6%, and Arena has since put Opus 5.5 above it on WebDev, leaving Astra only the image-to-WebDev board. It also roughly halves hallucination on AA-Omniscience, from Sol's 92% to 51% at max effort, and it still wins Zapier's AutomationBench and Terminal-Bench-Science 0.1 against Opus 5.5. You need a paid plan: Astra is the first model OpenAI has designated Critical for cybersecurity, so approved defenders in its Daybreak program got it first, with ChatGPT Business and Pro following on September 4 and Plus a few hours later, and nothing has been announced for the free tier. Our full GPT-6 Astra breakdown covers the benchmarks and the caveats. The family grew on September 22 with GPT-6 Sol at $2 / $10 and GPT-6 Luna at $0.10 / $0.50, which Artificial Analysis scores 47.5 and 37.3 against Astra's 52.7. Both are in ChatGPT Work, Codex and the API, and OpenAI says neither is in Chat yet; our GPT-6 Sol and Luna breakdown has the rest. OpenAI added GPT-6.1 Sol on September 29 at the same $2 / $10 as GPT-6 Sol, and Artificial Analysis scores it 51.8, 0.9 behind Astra at a fifth of the token price. There is no GPT-6.1 Astra: the Wall Street Journal reported that OpenAI cancelled its October release after internal safety tests.

What is Claude Opus 5?

Claude Opus 5 is Anthropic's July 24, 2026 flagship, and it was the #1 model on Artificial Analysis until Claude Fable 5.1 arrived on September 1. Claude Opus 5.5 replaced it on September 22, and Anthropic now lists Opus 5 as a legacy model. On Intelligence Index v4.3.2 it scores 50.8, behind Opus 5.5 at 57.6, Claude Sonnet 5.5 at 56.0, Claude Fable 5.1 at 53.4 and GPT-6 Astra at 52.7, and it is fourth among distinct models on both of Artificial Analysis's agentic boards, with 1673 on AA-Briefcase v1.1 and 1708 on GDPval-AA v2.1. API pricing is $5 / $25 per 1M tokens, the same as Opus 4.8 and more than Opus 5.5's $4 / $20, with a 1M-token context window and a May 2026 knowledge cutoff. ARC Prize independently confirmed Anthropic's launch claim on ARC-AGI-3, where Opus 5 scored 30% against 8% for the next-best model at the time, though GPT-6 Astra has since passed it. Arena places it fourth on WebDev, third on image-to-WebDev, first on Document and eleventh on text. One thing to keep in mind: Artificial Analysis measures its hallucination rate at 50%, which is why it holds no accuracy crown on this page.

What is Claude Fable 5.1?

Claude Fable 5.1 is Anthropic's September 1, 2026 release, and it was the highest-scoring model Artificial Analysis had measured until Claude Opus 5.5 arrived on September 22. It now scores 53.4 on Intelligence Index v4.3.2 at max effort, third among distinct models behind Opus 5.5 at 57.6 and Claude Sonnet 5.5 at 56.0. It still holds the highest raw factual accuracy on AA-Omniscience at 67%, which keeps this page's accuracy crown, and the best combined placing on the three prose boards, which keeps the writing crown. It shipped alongside Claude Mythos 5.1, which is the same model with the biology and cybersecurity safeguards relaxed, available by invitation only and with no published eligibility criteria. Pricing is $10 / $50 per 1M tokens, unchanged from Fable 5, but cache reads fall 75% to $0.25, on a 1M-token context window with a June 2026 knowledge cutoff. It is included in Claude Max and Team Premium at roughly 50% of weekly limits and reachable from Pro through usage credits. For most workloads Claude Opus 5.5 is now the better buy, because it scores higher at less than half the token price. Our full Claude Fable 5.1 and Mythos 5.1 breakdown covers the system card and the safeguard split.

Is Claude Fable 5 back?

Yes. Anthropic redeployed Claude Fable 5 on July 1, 2026 after the US government lifted the export-control restriction it had imposed on June 12. It is available again on the Claude API, Claude.ai, Claude Code, and Claude Cowork. Anthropic made Fable 5 permanent in the paid plans on July 20, 2026. Max and Team Premium include it at roughly 50% of regular usage limits; Pro and Team Standard reach it through usage credits with a one-time $100 starting credit. API pricing is $10 / $50 per million tokens. Fable 5 is a Mythos-class model built for long-horizon agentic work with a 1M-token context, and it is the runner-up for coding behind Claude Opus 5, at #2 on Arena's image-to-WebDev board (1,625.7) and #4 on WebDev.

What is GPT-5.6 and can I use it?

Yes, as of July 9, 2026. GPT-5.6 is OpenAI's next-generation model family, and after a two-week gated preview that began June 26 behind a US-government safety review, it reached general availability on July 9. There are three tiers, least to most capable: Luna, the fast, cheapest tier ($0.20 / $1.20 per 1M tokens); Terra, a balanced everyday model OpenAI says matches GPT-5.5 ($2 / $12); and Sol, the flagship tuned for biology, chemistry, and cybersecurity ($4 / $20). OpenAI cut Luna by 80% and Terra by 20% on July 30, 2026 and left Sol alone, then cut Sol by more than 20% on August 21, 2026 as a promotional rate for about three months. No ChatGPT subscription price has changed. It is now live across ChatGPT, Codex, and the API as OpenAI's default model. One caveat: OpenAI's system card and the external evaluator METR flagged elevated "scheming" behaviour in Sol, so treat it carefully for high-stakes factual work until independent results settle. One naming point since September 22: OpenAI reused Sol and Luna for the GPT-6 generation, so an unqualified Sol or Luna is now ambiguous. GPT-6 Sol is $2 / $10 and GPT-6 Luna $0.10 / $0.50, against $4 / $20 and $0.20 / $1.20 for the 5.6 tiers, which OpenAI still sells.

Is Grok 4.7 out yet?

Yes. SpaceXAI released Grok 4.7 on September 21, 2026, replacing Grok 4.6 as the flagship. Unlike 4.6, which was a post-training refresh of the same base, 4.7 uses a new and larger base model, trained with a longer reinforcement learning run weighted toward jobs that take many hours; SpaceXAI publishes no parameter count. Artificial Analysis scores it 46.3 on Intelligence Index v4.3.2 at high effort against Grok 4.6's 44.3, and it places above GPT-6 Astra on both agentic boards, 1695 to 1542 on GDPval-AA v2.1 and 1657 to 1569 on AA-Briefcase v1.1. Pricing is unchanged at $2 / $6 per 1M tokens on a 500K-token context window, though prompts of 200,000 tokens and above re-bill the whole request at $4 / $12, and a fast variant runs at twice the output speed for twice the price. The cost per Intelligence Index task did move, from $1.86 to $2.73, because 4.7 emits roughly twice the output tokens. It is live in the Grok API, Cursor, Grok Build, third-party coding harnesses and cloud routers. SpaceXAI's announcement says nothing about the consumer Grok app, which still serves Grok 4.6 at $30/month. Grok is also our creativity pick, on product grounds rather than prose quality; for the best-written output, Claude Fable 5 wins outright. Full numbers are in our Grok 4.7 breakdown.

Which AI is the best for coding?

Claude Opus 5.5 is the best for coding, and its cheaper sibling is close behind. Opus 5.5, released September 22, leads Arena's WebDev board at 1820 against GPT-6 Astra's 1792 and scores 71.7 on LiveBench Agentic Coding, second only to DeepSeek V4.1-Flash and 15 points above Claude Sonnet 5.5, at $4 / $20 per 1M tokens. Sonnet 5.5, released September 28 at $2 / $10, leads Artificial Analysis's Terminal-Bench 4.0 at 63.6% against Opus 5.5's 59.6% and LiveBench Coding at 91.4 against 89.3, which makes it the value pick as long as you keep it below max effort, where an index task costs $7.60 against Opus 5.5's $5.98. On Anthropic's own harnesses Opus 5.5 is ahead on FrontierCode v1.1, 54.4% to 52.1%, and on CursorBench 4.0, 57.8% to 55.5%. GPT-6 Astra is third and keeps Arena's image-to-WebDev board at 1733. Among open weights, GLM-5.3 is strongest on Terminal-Bench 4.0 at 41.9% and Kimi K3 is the highest-placed open model on Arena WebDev at tenth, while GLM 5.3 Flash is the one most teams can host, at $0.25 an index task under plain MIT. Artificial Analysis has retired both its Coding Index and its Coding Agent Index.

Which AI is the best for writing?

Claude Fable 5.1 is the best for writing, because it has the best combined placing across the three prose boards: second on EQ-Bench Creative Writing v3, third on LiveBench Language and seventh on Arena's creative-writing board. Claude Opus 5.5 is the pick if you trust Arena's human votes most: it now leads Arena's creative-writing board at 1521, though it is only seventh on EQ-Bench and eighth on LiveBench Language. Claude Fable 5 tops LiveBench Language at 90.7 and is second on Arena's creative-writing board, but has slipped to eleventh on EQ-Bench. Fable 5.1 and Fable 5 cost $10 / $50 per 1M tokens. Claude Sonnet 5.5 is the value pick and what most people should actually use, since it is free on claude.ai and $2 / $10 in the API, though no creative-writing board has rated it yet. For professional deliverables, Claude Opus 5.5 leads GDPval-AA v2.1 at 1846 with Sonnet 5.5 two points behind at 1844. Kimi K3 is fifth on EQ-Bench at 2082.3 if you want distinctive fiction, GPT-5.5 is the alternative for fact-anchored business writing, and Gemini 3.8 Flash is the price-performance pick for bulk content.

What is the best open-weight AI model in 2026?

Xiaomi's MiMo-V2.6-Pro, published on September 21, 2026 under plain MIT. Artificial Analysis measures it at 46.3 on Intelligence Index v4.3.2, the highest open score it has ever published, above GLM-5.3 at 44.8 and Kimi K3 at 43.6, and it runs an index task for $0.13, cheaper than any other open model on this page. It is a 1.02-trillion-parameter mixture of experts with 42 billion active, a 1M-token context and the weights on Hugging Face from day one. Two caveats: Arena's human voters rate it only 21st on WebDev and 23rd on text, and it has barely a week of independent track record. Kimi K3 remains the highest-placed open model on Arena WebDev, tenth, while Tencent's Hy4 Preview is now the highest open model on the Agent Arena at twelfth, and GLM-5.3 is still the strongest open model on Artificial Analysis's Terminal-Bench 4.0 at 41.9%, on a bespoke Z.ai licence rather than MIT. For teams that actually want to host their own, GLM 5.3 Flash (Z.ai, MIT) remains the practical recommendation at 41.8 on the index and 320B total with 18B active, because a 1.02-trillion-parameter model needs a cluster and the Kimi K3 download is 96 shards and about 1.56 TB.

What is the cheapest frontier-class AI model?

On API pricing per million tokens, GPT-6 Luna is the cheapest closed frontier-class model at $0.10 / $0.50, after OpenAI halved it on September 22, 2026, and Artificial Analysis scores it 37.3 on its v4.3.2 index against Gemini 3.6 Flash's 34.0. Among open weights the pick is still GLM 5.3 Flash, our price-performance pick since August, now at its plain $0.15 / $0.50 list since the launch promotion expired on September 9, scoring 41.8 under a plain MIT licence you can self-host. DeepSeek held that pick until it raised rates roughly fourfold on August 16, and it came close to taking it back on September 10, when V4.1-Flash replaced V4-Flash at $0.15 / $0.60 off-peak and $0.30 / $1.20 at peak: off-peak the two now tie on input, with GLM cheaper on output. Artificial Analysis still separates them on cost per index task, $0.25 for GLM against $0.27 for DeepSeek. Google's cheap tier improved again on September 2, with Gemini 3.8 Flash scoring 40.9 at the same introductory $0.75 / $3.75, though that rate doubles on January 1, 2027. MiniMax M3 lists at $0.30 per million input tokens on a permanent 50% discount, with LongCat-2.0 as a strong MIT alternative. Measured per Intelligence Index task rather than per token, the cheapest model on this page is GPT-6 Luna at $0.07, ahead of GPT-5.6 Luna at $0.18 and the open-weight MiMo-V2.6-Pro at $0.13.

Which AI models are free?

ChatGPT Free now defaults to GPT-5.6 (with GPT-5.5 still available) under usage limits. Free and Go users can also reach GPT-6 Luna in the ChatGPT desktop app since September 22. Gemini Free runs Gemini 3.6 Flash in the Gemini app and Google AI Studio. Claude Free runs Claude Sonnet 5.5, released September 28, with usage limits. DeepSeek Chat runs DeepSeek V4 free on the DeepSeek website. Grok has a limited free consumer plan (X Premium is a paid add-on). Qwen 3.5, Qwen3.8-Flash-Next, NVIDIA Nemotron 3 Ultra, MiniMax M3, LongCat-2.0, Kimi K2.6, DeepSeek V4, GLM-5.2, GLM 5.3, GLM 5.3 Flash, MiMo-V2.6-Pro and Tencent's Hy4 Preview are open-weight and free to self-host, though the licences differ and only some of them are plain MIT or Apache. Qwen 3.7 Max is API-only with no consumer chat front-end, but Alibaba now includes 200 free model requests per day. Kimi has a free basic tier in its app, with heavier agentic use metered.

What is Gemini Spark and which plan do you need?

Gemini Spark is Google's first 24/7 cloud-resident AI agent, launched at Google I/O on May 19, 2026 and no longer an Ultra exclusive: since July 30, 2026 it also reaches the $19.99/month Google AI Pro tier, in the US and more than 160 further countries, while Google AI Ultra at $99.99 or $199.99/month carries the wider global rollout. Google AI Plus, the free tier and work or school accounts do not include it. Spark is built on Gemini base models with Google's Antigravity harness on a Google Cloud VM, integrates with Gmail, Google Docs, and other Google Workspace apps, and can interact with Chrome and Android's Halo system on the device side. It is worth the spend for users who have repeatable long-running workflows (inbox triage, research roll-ups, scheduled tasks). For one-off tasks, Claude Cowork at $20/month covers most desktop-agent needs.

What is Fello AI?

Fello AI is an AI chatbot for Mac, iPhone, and iPad that lets you use all top AI models like ChatGPT, Claude, Gemini, Grok, and DeepSeek in one app, with models updated regularly so you always have the latest. It is $9.99/month with a 4.7-star rating across 27,000+ reviews.

How often do you update this page?

We update this page at least monthly and within 24-48 hours of any major model launch.