Sakana AI Releases Fugu Max and Fugu Ultra v2, Models That Orchestrate Other Models Instead of Being One

On 11 September 2026, Sakana AI announced two new releases, Fugu Max and Fugu Ultra v2, that do not add a bigger language model to the field so much as remove the assumption that one big model has to do all the work. Both are orchestrator systems: they take a task, break it up, and dispatch the pieces to a pool of smaller open-weight and specialist models, then assemble the results. Sakana AI, a Tokyo-based lab, describes the pair as "the same core orchestration architecture optimized for two distinct missions" — Fugu Max for cost efficiency, Fugu Ultra v2 for maximum output quality on hard, multi-step problems.

What actually changed

This is the third public iteration of an idea Sakana AI has been building since opening a beta of the original Fugu system in April 2026, ahead of a general release that June. That first version was pitched against a specific inefficiency: in the company's words, "conventional approaches to utilizing foundation models often require users to manage multiple API keys, as models from different providers tend to specialize in distinct areas." In the September release announcement, Sakana AI frames the same problem more bluntly, arguing that "a system that deploys a multi-trillion-parameter model to execute a simple data lookup is not intelligent, but wasteful." Instead of one model doing everything, a Fugu-trained router decides which of several available models — including, per the September release, models from NVIDIA's Nemotron line — should handle which part of the job, in what order, and whether the result needs a second pass.

The architecture and training method behind this are documented in a technical report Sakana AI's research team posted to arXiv in June 2026, which describes combining large-scale fine-tuning, evolutionary search and reinforcement learning to teach a model to construct agent scaffolds on the fly rather than follow fixed routing rules.

How the routing works

According to Sakana AI's own description, the system "learns to dynamically assemble agents from a pool and coordinate them through non-obvious but highly efficient collaboration patterns," handling the collaboration topology, role assignment and subtask dispatch itself rather than through hand-written logic. A separate capability the company highlights is recursive self-calling: when permitted, the orchestrator can invoke another instance of itself to check or extend a first answer, which Sakana AI frames as "a new form of test-time scaling" where the model "recognizes when its first attempt fell short and spins up a corrective workflow." The recursion depth can be adjusted at inference time without retraining the underlying model.

The benchmark numbers

Sakana AI reported that Fugu Ultra v2 scored 48.3 on Chartography, a visual reasoning and data-interpretation benchmark, in its September 2026 release, compared with 27.3 for Anthropic's Opus 5 and 29.5 for its Fable 5 on the same test. On DeepSWE, a real-world software-engineering benchmark, Fugu Ultra v2 scored 74.3, which Sakana AI says places it ahead of models priced several times higher per token. Across the eight agentic benchmarks Sakana AI uses for evaluation, it reports Fugu Ultra v2 placing best or joint-best on five and in the top two on seven, while Fugu Max separately takes the best overall score on six benchmarks, including Terminal Bench 2.1, GPQA Diamond and AA-LCR. The company states plainly that "Fugu Ultra v2 achieves these scores without Fable 5, Fable 5.1, or GPT-6-Astra in its agent pool" — the point being that the orchestration layer, not access to any single proprietary frontier model, is producing the result.

On pricing, Sakana AI's console lists Fugu Max at a flat $2 per million input tokens and $6 per million output tokens regardless of context length, while Fugu Ultra v2 costs $5 per million input tokens and $30 per million output tokens for context up to 272,000 tokens, rising to $10 and $45 per million respectively beyond that threshold.

What an independent read adds

An independent analysis published by Pondero the following day flagged two caveats worth keeping in view. First, Fugu Max's headline comparisons are against similarly priced models, not the current frontier leaders, so the efficiency claim is narrower than it first appears. Second, the strongest results cluster in specific domains — visual reasoning and coding — and Sakana AI's own release does not claim the same margin generalizes to other task types. Pondero's analysis also noted that Sakana AI's pricing above the 272,000-token threshold was, at the time of its writing, still unclear; that detail has since been published on Sakana AI's console pricing page, cited above.

Why it matters

For engineering teams building agentic pipelines, the significant claim here is not a benchmark score but the architectural bet: that coordinating several smaller, swappable, often open-weight models can match or beat a single proprietary frontier model on real tasks, at a fraction of the per-token cost, and without being locked to any one provider's roadmap. Whether that holds outside Sakana AI's own benchmark suite is something outside labs and independent evaluators will need to test directly, ideally by running the same tasks against Fugu Max and Fugu Ultra v2 under conditions the vendor did not choose.