Qwen3.7-Max Review: Benchmarks, Pricing and How It Compares to Qwen3.8-Max

Alibaba’s Qwen3.7-Max launched on May 20, 2026 at the Alibaba Cloud Summit in Hangzhou, and the numbers are loud. It scored 56.6 on the Artificial Analysis Intelligence Index at launch (good for #5 that week, at the time top 10 of 151 measured models), the highest-ranked Chinese AI model on that leaderboard to date. It carries a 1 million-token context window, costs $2.50 per 1M input tokens, and Alibaba’s internal testing reports a 35-hour autonomous coding run that fired 1,158 tool calls and hit a 10× speedup over the standard Triton reference. Since then, Alibaba has replaced it at the top of the lineup with Qwen3.8-Max, generally available on August 3, 2026 at a lower $2 / $6. Artificial Analysis now marks Qwen3.7-Max as deprecated and no longer scores it on its v5.0 index.

At launch the question was whether it beat GPT-5.5, Claude Opus 4.7, or Gemini 3.5. Short answer, on a price-per-intelligence basis, Qwen3.7-Max was one of the strongest frontier models you could call through an API. Long answer, there’s a caveat in how it earns its hallucination score that you’ll want to know about before betting your workflow on it. We tested the released benchmarks, pulled the official pricing, and broke down what this drop actually means for the coding, agent, and AI tooling space heading into mid-2026.

The Key Takeaways

• Qwen3.7-Max scored 56.6 on the Artificial Analysis Intelligence Index at launch, top 10 globally and then the highest placement for a Chinese model
• 1 million-token context window and a max output of 65,536 tokens, with extended thinking enabled by default
• API pricing is $2.50 input / $7.50 output per 1M tokens, with cached input at $0.25 per 1M on explicit caching (90% off) or $0.50 on the automatic implicit cache
• Demonstrated 35 hours of continuous autonomous coding, 1,158 tool calls, and a 10× geometric speedup vs the Triton reference kernel
• Proprietary model (no open weights), available via Alibaba Cloud Model Studio, OpenRouter, and direct Claude Code integration through the Anthropic API protocol

What Is Qwen3.7-Max?

Qwen3.7-Max is Alibaba’s flagship proprietary reasoning model, designed as an “agent foundation” rather than a general chat assistant. It’s the successor to Qwen3.6 Max Preview, and the first Qwen model to break the 1M-token context barrier, jumping up from 256K. It uses extended chain-of-thought reasoning by default and is positioned to compete head-on with GPT-5.5, Claude Opus 4.7, and Gemini 3.5 Flash on tasks that need long-horizon autonomy.

Alongside Qwen3.7-Max, Alibaba quietly shipped Qwen3.7-Plus-Preview, a multimodal variant that adds vision input and runs at a lower price point. Plus is the budget pick for high-volume, routine workloads, while Max is the heavy hitter for reasoning, agentic coding, and document-scale context. Unlike many earlier Qwen releases, neither model is open-weight; both run only through Alibaba’s hosted API. In June 2026, Rio de Janeiro’s city government released Rio 3.5 Open 397B, an open-weight fine-tune that benchmarks itself against Qwen 3.7 Plus.

Benchmarks: How Qwen3.7-Max Stacks Up

The launch numbers are strong across the board. According to Artificial Analysis, Qwen3.7-Max sat between Gemini 3.5 Flash and Claude Opus 4.7 on the overall Intelligence Index at launch, but it punched above its weight on a handful of demanding reasoning benchmarks.

On individual benchmarks, Qwen3.7-Max posts 92.4 on GPQA Diamond (ahead of Claude Opus 4.6 Max at 91.3, behind GPT-5.5 at 93.6) and 97.1 on HMMT 2026 February, the highest score in its comparison group. It also lands 44.5 on Apex (well ahead of DeepSeek V4 Pro at 38.3), 80.4 on SWE-Verified for software engineering tasks, and 41.4 on Humanity’s Last Exam, narrowly beating Opus 4.6 Max at 40.0.

The honest caveat: Qwen3.7-Max’s low hallucination rate is partly an artifact of higher abstention. According to Officechai’s benchmark analysis, the model’s attempt rate fell to 48.0%, the lowest among comparable frontier models. In plain English, it refuses to answer more often, which lowers wrong answers but also lowers usefulness on edge cases. That trade-off matters if you’re plugging it into an agent that needs to push through ambiguity.

Pricing: $2.50 Input, $7.50 Output, $0.25 Cached

At launch, Qwen3.7-Max was one of the better-priced frontier models on the market, and Alibaba still lists it at the same rates. The API costs $2.50 per 1 million input tokens and $7.50 per 1 million output tokens, with a max output of 65,536 tokens per request. With explicit caching, cached input drops to $0.25 per 1M tokens, a 90% discount that makes repeated long-context calls cheap; the automatic implicit cache bills hits at $0.50, an 80% discount.

$0.25 cached input is the killer feature

For agent workflows that re-read the same codebase, document, or chat history across hundreds of turns, the cache discount turns Qwen3.7-Max into a pricing outlier. At launch, Claude Opus 4.7 charged $5 per 1M input tokens and $25 output, double the input and over triple the output rate of Qwen3.7-Max. GPT-5.5 was $5 input and $30 output, the most expensive of the bunch. Gemini 3.5 Flash ($1.50/$9) and DeepSeek V4 Pro ($1.74/$3.48) were cheaper but scored lower on reasoning benchmarks. If your workload is “long context, lots of tool calls, occasional thinking,” Qwen3.7-Max sat in the sweet spot between frontier intelligence and frontier pricing. Today its own successor, Qwen3.8-Max, undercuts it at $2 / $6.

Where you can pay for it

At launch the API went live on Alibaba Cloud Model Studio, OpenRouter, and Together AI, with Qwen Chat (chat.qwen.ai) offering limited free preview access through the web, and Model Studio and OpenRouter still list it. There is no permanent free tier, and the Plus-Preview variant is text-only on the API even though it accepts image input through the chat interface.

The 35-Hour Autonomous Run, Explained

The headline demo from launch is the 35-hour continuous autonomous coding run, and it deserves a closer look because most launch articles glossed over what actually happened. According to Alibaba’s internal testing (no independent verification has been published yet), Qwen3.7-Max was given an isolated server equipped with a Zhenwu M890 AI accelerator, a brand-new hardware architecture the model had never seen in training. The task was to optimize an attention kernel from scratch.

Over 35 straight hours, the model executed 1,158 distinct tool calls, ran 432 kernel evaluations, diagnosed compilation failures on its own, and iteratively rewrote its code. It ended with a 10× geometric mean speedup over the Triton reference implementation, per VentureBeat. That single demo isn’t proof of general agent intelligence, but it does show the model can hold context, recover from failure, and stay coherent across a length where most other reasoning models start hallucinating variables or looping.

For developers, the practical takeaway is this. If you’re running a coding agent or a research assistant overnight, Qwen3.7-Max was, at launch, one of the few models tested at this duration. Whether your wallet can keep up with the token spend is a separate question.

Where to Try Qwen3.7-Max

There are four paths to get hands-on, and your choice depends on whether you want a chat interface, an API, or an agent harness.

Qwen Chat (chat.qwen.ai) is the free preview entry point. At launch it had a daily message limit and exposed both Qwen3.7-Max and Qwen3.7-Plus-Preview through a dropdown. Useful for kicking the tires before committing to API spend.

Alibaba Cloud Model Studio is the official API home with full pricing applied and the longest context limits. You’ll need an Alibaba Cloud account, which is more friction than other providers but enables prompt caching and direct support.

OpenRouter listed the model at the same price tier at launch and lists it lower today, at $1.475 / $4.425 when read on October 30, 2026; it is also easier to sign up for, particularly if you’re already routing multiple models through a single key. Together AI also hosted it at launch for fine-tuning-friendly workloads.

Claude Code is the one that surprised people. Qwen3.7-Max natively supports the Anthropic API protocol, which means you can point Claude Code, OpenClaw, or any other Anthropic-compatible harness at the Qwen endpoint and run it as a drop-in. If you’ve been priced out of Claude Code pricing but want the same workflow, this is the cheapest credible alternative you’ve had access to in months.

If you’d rather skip the API setup and just compare frontier models from one app, the Fello AI Mac and iOS app bundles Claude, ChatGPT, Gemini, Grok, DeepSeek and Qwen 3.8 Max, the successor to Qwen3.7-Max, under one subscription from $11.99 a month, so you can A/B test prompts across providers without juggling separate billing accounts. On a Mac that includes Qwen as a native desktop client, routed through US-based infrastructure rather than Alibaba Cloud directly.

Verdict: Should You Switch?

At launch, Qwen3.7-Max was the most credible Chinese frontier model yet, and the pricing made it the natural pick for cost-sensitive agentic and long-context work. If your bottleneck was GPT-5.5 or Claude Opus 4.7 API bills, it was the most direct way to cut spend without dropping into the “Gemini 3.5 Flash, but slightly worse at reasoning” tier.

You should switch if you’re building agents that run for hours, you’re handling million-token contexts, or you need the $0.25 cached input rate to make a use case profitable. You should not switch if your workload depends on full multimodal vision (use Plus-Preview or wait for Max-Vision), or if your application can’t tolerate the higher abstention rate on hard questions. For pure raw intelligence and instruction-following at the absolute frontier, GPT-5.5 still led on the Intelligence Index at launch. If you are choosing today, start with Qwen3.8-Max, which costs less and has replaced Qwen3.7-Max as Alibaba’s flagship.

The bigger story is what this launch signals. Alibaba is now shipping a proprietary model that ties directly to its own custom AI accelerator (the Zhenwu M890) and rack-scale server (Panjiu AL128), in the same week, on the same stage. That’s the full-stack AI play that NVIDIA and OpenAI are also racing toward, and Alibaba is the first Chinese vendor to credibly join that race. Qwen3.7-Max is the model layer of that bet, and on the numbers we have, the bet is paying off.

For the wider picture on how Chinese AI labs caught up this fast, our deep dive on the Chinese AI race and our piece on Kimi K2.5 give you the prior context. If you’re shopping for best free coding models, Qwen3.7-Max isn’t free, but it’s worth comparing on benchmark-per-dollar. For Moonshot’s answer in the same lineage, see our guide to Kimi K2.7 Code, its open-weight coding model.

Qwen wasn’t the only Chinese frontier move that month. MiniMax previewed MiniMax M3, a sparse-attention LLM claiming 9.7× and 15.6× speedups at the 1M-token mark over M2.5, signaling that efficiency at long context, not just raw scale, is the new battleground for Chinese labs.

FAQ