What Sakana released
Sakana AI has released Fugu Max and Fugu Ultra v2, two new versions of its Fugu system, which spreads each request across a pool of other AI models instead of answering from a single one. Both are available now through Sakana’s OpenAI-compatible API, according to the company’s release post, which is dated 11 September. MarkTechPost reported the launch on 10 September.
Fugu Max is the cheaper of the two. It costs $2 per million input tokens and $6 per million output tokens, and Sakana says its output pricing is 40–60% below Sonnet 5, GPT 5.6 Terra and Kimi K3. Fugu Ultra v2 is the quality-first tier, aimed at harder reasoning and software work.
A system sold as a model
Sakana describes Fugu as a “multi-agent system, delivered as one model”. A developer calls one endpoint. Behind it, a learned orchestrator decides which models from the pool to assemble for the task, which roles they play and how they coordinate, rather than running a fixed workflow. The company points to two of its ICLR 2026 papers as the basis: TRINITY, in which a lightweight evolved coordinator assigns thinker, worker and verifier roles across turns, and Conductor, trained with reinforcement learning to find coordination strategies expressed in natural language.
The release post says the goal is to route each task to the leanest model capable of solving it. Sakana says the pool now includes a large number of open-weight and specialised models, among them Nvidia’s Nemotron family through a collaboration with Nvidia. It does not disclose which models answer a given query.
The benchmark claims
Every figure below comes from Sakana, and none has been independently reproduced.
For Fugu Max, Sakana reports the best overall score on six benchmarks: Terminal Bench 2.1, GPQA Diamond, AA-LCR, GDP.pdf, AutomationBench and SWEFish. SWEFish is, in Sakana’s own description, an internal benchmark built from the company’s own coding challenges, so outsiders cannot check it. Sakana also says Fugu Max extends the cost-performance frontier on seven of the ten benchmarks it tested.
For Fugu Ultra v2, Sakana reports the best or joint-best score on five of eight benchmarks (GDP.pdf, Chartography, SWEFish, DeepSWE and Toolathon) and a top-two place on seven of eight. The widest reported gap is on Chartography, a visual reasoning and data-interpretation test, where Sakana gives Fugu Ultra v2 48.3 against 27.3 for Opus 5 and 29.5 for Fable 5. On DeepSWE, a real-world software engineering benchmark, it reports 74.3.
Sakana also stresses what is not in the pool. It says Fugu Ultra v2 reached these scores without Fable 5, Fable 5.1 or GPT-6 Astra among its agents. The argument is that coordinating other models can compete with frontier models head-on. The flip side is that Fugu’s quality depends on which third-party models Sakana is able to put in the pool.
Pricing and access
The Fugu product page lists four tiers: the base Fugu model for everyday coding; Fugu Ultra at $5 per million input tokens and $30 per million output ($10 and $45 above 272,000 tokens of context); Fugu Max at $2 and $6; and Fugu Cyber, a security-focused version priced through sales. Subscription plans run from $20 to $200 a month, and token use and cost are reported on every request.
European developers cannot use it yet. Sakana’s product page says Fugu is not available in the EU or EEA while the company works towards compliance with GDPR and EU-specific regulation. Elsewhere it is offered globally, though Sakana says access may vary with local conditions.
What to watch
The claims that matter most are about cost, and those are the easiest to test. Independent evaluators can price and score Fugu Max against Sonnet 5, GPT 5.6 Terra and Kimi K3 on public benchmarks, which SWEFish is not. The other thing to watch is the EU and EEA restriction, and whether Sakana gives a date for lifting it.
