Sakana Fugu Ultra: Japan's Premium 1M-Context Model

Sakana AI just shipped a model that is not really a model. Fugu Ultra, launched June 22, 2026 by the Tokyo lab and live on CoreAI since June 24, is a learned multi-agent orchestration system wearing a single API endpoint: send it a request, and a trained coordinator decides whether to answer directly or assemble a team of specialist models to attack the problem together. The result posts numbers that turned heads — 73.7% on SWE-Bench Pro, ahead of every single-model competitor Sakana tested — and a price tag that makes clear who it is for. Sakana Fugu Ultra is Japan's most ambitious entry yet into the frontier tier, and it arrives with a genuinely different architecture story.

What Exactly Did Sakana AI Launch?

Sakana AI has built its reputation on nature-inspired, unconventional research — and Fugu is the productized version of two of its ICLR 2026 papers: TRINITY, an evolved LLM coordinator, and Conductor, which taught models to orchestrate other agents in natural language. Instead of scaling one giant network, Sakana trained a language model whose specialty is routing: reading your request, deciding which expert models should handle which pieces, coordinating their work, and merging the result into one answer. The official Fugu page calls it a "multi-agent system as a model," which is unusually honest product naming.

The family launched in two flavors. Fugu handles everyday, latency-sensitive work with a shallower pool of helpers. Fugu Ultra is the maximum-quality variant: on hard, multi-step problems it convenes a deeper bench of expert agents and spends more compute coordinating them. From the outside, both behave like any other chat model — the orchestration happens entirely behind the endpoint, as MarkTechPost's launch coverage details. One consequence worth understanding: because the pool is swappable, some subtasks may be handled by third-party frontier models under the hood. That is the design, not a leak — but teams with strict data-routing requirements should read Sakana's documentation before sending sensitive material.

How Strong Is the 73.7% SWE-Bench Pro Claim?

The benchmark table is the launch's loudest argument. On SWE-Bench Pro — the harder, repository-level successor to SWE-Bench Verified — Sakana reports Fugu Ultra at 73.7%, with Claude Opus 4.8 at 69.2%, GPT-5.5 at 58.6%, and Gemini 3.1 Pro at 54.2%. If a team-of-models can reliably outscore the best single models, that is a meaningful data point in the scaling debate: coordination, not just parameter count, as a path up the capability curve.

The caveats are the standard ones, plus one novel wrinkle. Standard: these are vendor-reported numbers on a benchmark the vendor chose. Novel: when an orchestrator routes work to a pool that can include other labs' frontier models, a benchmark win is partly a measurement of the pool, not just the coordinator. That does not make the score wrong — you buy the system's output, not its org chart — but it reframes what is being measured. The practical test, as ever, is your own workload: run Fugu Ultra head-to-head against Opus 4.8 or Claude Sonnet 5 in CoreAI Compare on a task you actually care about and see which answer you would ship.

What Does Fugu Ultra Cost, and What Do You Get?

Premium architecture, premium bill. Here is how Fugu Ultra's pricing and specs line up against the frontier models it is gunning for, per CoreAI's live catalog:

Note the tiered long-context pricing: past 272,000 prompt tokens, rates double on input and rise 50% on output — coordinating a model team across a truly full million-token window is expensive compute, and Sakana passes that through. Cached input starts at $0.50 per million. The spec sheet earns some of that money back: a 1M-token window, 128K output cap, image input, reasoning, and native web search support in one package. For how the million-token club compares more broadly, our guide to 1M-token context models covers the field Fugu Ultra just joined.

So who should pay it? Fugu Ultra makes sense for the hardest 5% of your queue: multi-step research synthesis, repository-scale engineering decisions, long-document analysis where a wrong answer costs more than $30 worth of tokens. Routing your everyday questions to it is like hiring a consulting firm to fix a typo. The sensible pattern — easy to run on CoreAI, where one subscription spans all 300+ models — is a cheap default model for volume and Fugu Ultra reserved for the problems that deserve a committee.

Why does the Japan angle matter?

Fugu Ultra is also a milestone for Japan's AI ecosystem, which has watched the frontier race play out largely between American and Chinese labs. Sakana's bet sidesteps the compute arms race entirely: instead of out-training a 2-trillion-parameter monolith, train a smart coordinator and let the pool provide the horsepower. It is a strategy available to labs without the largest GPU fleets — and if the benchmark results hold up under independent testing, expect more orchestration-first releases from labs in the same position. The pufferfish naming is fitting: a small creature that becomes formidable by clever inflation. Whether the approach scales beyond benchmarks into daily production use is the question the next few months will answer, and it is one of the more interesting open questions in AI right now.

Frequently Asked Questions

What is Sakana Fugu Ultra?

It is the high-performance variant of Sakana AI's Fugu family, launched June 22, 2026. Rather than a single neural network, it is a trained multi-agent orchestration system: a coordinator model that routes your request across a pool of specialist models and merges their work into one answer, delivered through a single endpoint.

How is Fugu Ultra different from the standard Fugu model?

Fugu is tuned for everyday, latency-sensitive tasks with lighter orchestration. Fugu Ultra targets maximum answer quality on hard, multi-step problems, coordinating a deeper pool of expert agents — at higher cost and latency. Same interface, different depth of deliberation.

What are Fugu Ultra's context window and pricing?

A 1,000,000-token context window with up to 128,000 output tokens per response. Pricing is $5 per million input tokens and $30 per million output for standard requests, stepping up to $10 and $45 once a prompt exceeds 272,000 tokens. It is a premium-tier model on CoreAI.

Did Fugu Ultra really beat Claude Opus 4.8 on benchmarks?

On Sakana's reported SWE-Bench Pro results, yes: 73.7% versus 69.2% for Opus 4.8, with GPT-5.5 and Gemini 3.1 Pro further behind. The numbers are vendor-reported, and because Fugu's orchestration can route work across a pool of models, the score reflects the whole system rather than a single network. Independent verification is worth watching for.

Can I try Sakana Fugu Ultra without an API key?

Yes. Fugu Ultra is live on CoreAI across iOS, Android, and the web — one subscription covers it along with 300+ other models, with side-by-side comparison, file attachments, and web search built in. No API keys or per-token billing to manage.