Sakana AI has released Fugu-Ultra v1.1, an upgrade to its frontier orchestration model that the lab says gains up to 7.9 points over v1.0 across every benchmark it published, all at the same price. The Japanese company credits the jump to swapping the latest frontier models into the pool that Fugu-Ultra commands, with the strongest gains on coding and terminal tasks.
The update matters because Fugu-Ultra is not a single model, it is a system that routes each part of your request to the best frontier LLM in a swappable pool, including Claude Opus 4.8, GPT-5.5, and Gemini 3.1 Pro, then combines their answers behind one API. Those underlying models keep advancing, with Claude Opus 5 now topping the Artificial Analysis leaderboard. That design means Fugu-Ultra improves every time the models it orchestrates improve, and v1.1 is the first release to show that loop in action. Below you get the verified v1.1 benchmark scores, an honest head-to-head against the frontier models it beats, the full pricing, and who should actually use it.
The Key Takeaways
- Fugu-Ultra v1.1 improves on v1.0 by up to 7.9 points, according to Sakana, with the biggest gains on coding and agentic tasks.
- It scores 82.1% on Terminal-Bench 2.1 and 73.7% on SWE-Bench Pro, leading Sakana’s reported comparison against Opus 4.8, GPT-5.5, and Gemini 3.1 Pro.
- Pricing is unchanged: $5 per million input tokens and $30 per million output tokens.
- Fugu-Ultra is an orchestration model, a 7-billion-parameter conductor that routes work across a pool of frontier LLMs behind one OpenAI-compatible API.
- All scores are Sakana-reported, not independent, so read the leads as a vendor claim rather than a settled result.
What Is Fugu-Ultra?
Fugu-Ultra is Sakana AI’s quality-first model in the Fugu family, which launched on June 22, 2026. Unlike a normal chatbot, Fugu is trained to understand your request and build an agentic scaffold on the fly, delegating each sub-task to whichever frontier model handles it best. Sakana describes it as “a multi-agent system as a model,” and exposes all of it through a single API call. For the full picture of how the base system works, read our guide to Sakana Fugu.
The standard Fugu targets everyday, latency-sensitive work. Fugu-Ultra is tuned for the hardest problems, where answer quality matters more than speed. It suits AI research, academic paper reproduction, cybersecurity analysis, and large-scale patent investigations. Responses can take anywhere from 8 to 160 seconds, which rules it out for real-time chat.
Under the hood sits a roughly 7-billion-parameter orchestrator whose only job is to decide which external model should handle each step, then verify and synthesize the results. The approach is grounded in two Sakana research efforts, TRINITY, an evolved LLM coordinator, and Conductor, its work on teaching models to orchestrate agents in natural language. Routing decisions stay proprietary, so you cannot see which model answered a given request, and Anthropic’s Fable 5 and Mythos are absent from the pool because of export controls.
What’s New in Fugu-Ultra v1.1
The headline change is simple. Sakana refreshed the frontier models inside the pool, and because Fugu-Ultra’s job is to route work to the best available model, the whole system moves forward with them. The lab reports gains of up to 7.9 points over v1.0 across every benchmark it showed, though it has not published the underlying v1.0 scores for a line-by-line comparison, so treat the figure as Sakana’s own claim.
Those gains are not spread evenly. Sakana points to coding and terminal tasks as the strongest movers, the areas where agentic tools live or die on real developer work. That focus tracks with how early users described the product, one engineer said Fugu-Ultra surfaced more than twenty code-review issues where other tools flagged about three.
The best part for existing users is the price. Fugu-Ultra v1.1 costs exactly what v1.0 did, so the upgrade is a straight capability increase with no new bill. Sakana frames it plainly, the frontier keeps moving, and Fugu keeps getting better.
Fugu-Ultra v1.1 Benchmarks
Here are the verified v1.1 scores from Sakana’s official model page. They cover software engineering, agentic terminal use, competitive coding, and graduate-level reasoning. Every number is Sakana-reported using the company’s own scaffolding, not an independent lab result.
| Benchmark | What it tests | Fugu-Ultra v1.1 |
|---|---|---|
| SWE-Bench Pro | Real-world software fixes | 73.7% |
| Terminal-Bench 2.1 | Agentic terminal tasks | 82.1% |
| LiveCodeBench | Competitive coding | 93.2% |
| LiveCodeBench Pro | Harder competitive coding | 90.8% |
| GPQA-Diamond | Graduate-level reasoning | 95.5% |
| Humanity’s Last Exam | Broad expert knowledge | 50.0% |
The standout is Terminal-Bench 2.1, which measures whether a model can actually operate a terminal, run shell commands, manipulate files, and chain multi-step actions without a human babysitting it. That is the closest proxy to autonomous coding work, and it is where Fugu-Ultra’s orchestration approach pays off most.
How Fugu-Ultra Compares to GPT, Claude, and Gemini
Sakana benchmarks Fugu-Ultra v1.1 directly against the three frontier models it routes to. On the company’s numbers, Fugu leads across coding, terminal, and reasoning tests, though the margins tighten on reasoning.
| Benchmark | Fugu-Ultra v1.1 | Claude Opus 4.8 | GPT-5.5 | Gemini 3.1 Pro |
|---|---|---|---|---|
| SWE-Bench Pro | 73.7% | 69.2% | 58.6% | 54.2% |
| Terminal-Bench 2.1 | 82.1% | 74.6% | 78.2% | 70.3% |
| LiveCodeBench | 93.2% | 87.8% | 85.3% | 88.5% |
| GPQA-Diamond | 95.5% | 92.0% | 93.6% | 94.3% |
Read the table with two caveats. First, the scores are Sakana-reported, and vendor benchmarks use custom scaffolding and prompt tuning that flatter each model, so the gaps are not a clean, like-for-like result. Second, the lead is not a blowout everywhere; on Humanity’s Last Exam, Fugu-Ultra’s 50.0 barely edges Opus 4.8’s 49.8, a rounding error.
The deeper point is architectural. Fugu-Ultra does not beat these models by being a smarter single brain, it wins by orchestrating them, the same multi-agent approach other agentic tools use, and stitching their outputs into one result. Its ceiling is set by the same frontier models it competes against, then pushed higher by coordination, which is exactly why a model-pool refresh like v1.1 moves the numbers.
Fugu-Ultra Pricing
Pricing did not change with v1.1. Fugu-Ultra is billed per token through its API, with a higher rate that kicks in once your context passes 272,000 tokens.
| Token type | Standard rate | Above 272K context |
|---|---|---|
| Input | $5 / 1M tokens | $10 / 1M tokens |
| Output | $30 / 1M tokens | $45 / 1M tokens |
| Cached input | $0.50 / 1M tokens | $1.00 / 1M tokens |
If you prefer a flat subscription, Sakana offers three tiers, $20 per month for Standard, $100 per month for Pro, and $200 per month for Max. Access runs through an OpenAI-compatible endpoint, so you can point an existing OpenAI SDK at Fugu without rewriting your integration. At $5 input and $30 output, Fugu-Ultra sits among the pricier options, which fits its pitch as a tool for high-value work rather than casual chat.
Who Should Use Fugu-Ultra v1.1
Fugu-Ultra is aimed at hard, multi-step work rather than quick answers. Three use cases stood out during the beta and still define the product.
Code review and engineering. Fugu-Ultra digs deeper than single-model tools, catching issues others miss on real pull requests. The v1.1 coding gains make this its strongest pitch.
Security assessment. Cybersecurity engineers reported running an end-to-end analysis from a single instruction and getting back a clean report with evidence and retest steps.
Research and reasoning. Paper reproduction, Kaggle-style data science, and patent investigations all lean on the same orchestration strength, and the 95.5% GPQA-Diamond score backs the reasoning claim.
If you mostly want fast, conversational AI, the standard Fugu or a plain frontier model is the better fit, since Fugu-Ultra can take well over a minute per answer. If your work is agentic, long-running, and quality-sensitive, v1.1 is a meaningful step up at no extra cost.
Conclusion
Fugu-Ultra v1.1 is a clean win for anyone already using Sakana’s orchestration model, with up to 7.9 points of extra performance for the same price and the gains concentrated where agentic coding tools matter most. The release also proves the core idea behind Fugu, that an orchestration model improves automatically as the frontier moves, without a ground-up retrain. Just remember the benchmarks are Sakana’s own, so pilot it on your real workload before trusting the leaderboard. If you build agentic or coding workflows and want frontier-grade quality behind one API, start at Sakana AI’s Fugu page and point your existing OpenAI SDK at the endpoint.
FAQ
What is Fugu-Ultra v1.1?
Fugu-Ultra v1.1 is the upgraded version of Sakana AI’s high-end orchestration model. It commands a pool of frontier LLMs behind one API and, per Sakana, improves on v1.0 by up to 7.9 points across every published benchmark at the same price.
How much does Fugu-Ultra cost?
API pricing is $5 per million input tokens and $30 per million output tokens, doubling once context passes 272,000 tokens. Subscriptions run $20, $100, and $200 per month for Standard, Pro, and Max.
Is Fugu-Ultra better than GPT-5.5 or Claude Opus 4.8?
On Sakana’s own benchmarks, Fugu-Ultra v1.1 leads Opus 4.8, GPT-5.5, and Gemini 3.1 Pro on coding, terminal, and reasoning tests, scoring 82.1% on Terminal-Bench 2.1. The margins are thin on some reasoning tests, and the scores are vendor-reported, so results on your own work may vary.
Which models does Fugu-Ultra use?
Fugu-Ultra routes across a swappable pool of frontier LLMs including Claude Opus 4.8, GPT-5.5, and Gemini 3.1 Pro, and v1.1 refreshed that pool. Anthropic’s Fable 5 and Mythos are excluded because of export controls, and Sakana does not reveal which model answered a given request.
Do I need to change my code to use Fugu-Ultra?
No. Fugu-Ultra runs on an OpenAI-compatible endpoint, so you can point an existing OpenAI SDK at it without rewriting your integration.
