Sakana AI Ditches the Frontier Model for Fugu Ultra v2

Sakana AI shipped two models on September 11: Fugu Max, a cost-optimized option, and Fugu Ultra v2.0, a flagship system that isn't actually one model at all. It's a trained router that decides which model in a fixed pool, open-weight, specialized, or itself, should handle each piece of a task, and it never touches a proprietary frontier model to do it.

• Fugu Ultra v2.0 is a learned multi-agent orchestration system: a routing model that dispatches subtasks across open-weight and specialized models, and can call additional instances of itself recursively.
• It ships a 1,000,000-token context window, configurable reasoning effort (high, xhigh, max), function calling, structured outputs, image and PDF input, and built-in web search, all through one OpenAI-compatible API.
• Pricing lands at $5 per million input tokens and $30 per million output tokens.
• Sakana reports top or joint-top scores on five of eight of its own headline agentic benchmarks, a vendor-reported figure with no independent confirmation yet.

What did Sakana actually release?

Two models, both dated September 11. Fugu Max is positioned as the cheaper, faster option for everyday tasks. Fugu Ultra v2.0 is the one worth studying closely: Sakana describes it not as a bigger model but as a learned multi-agent orchestration system, delivered through a single OpenAI-compatible API endpoint at $5 per million input tokens and $30 per million output tokens. It carries a 1,000,000-token context window, three configurable reasoning-effort tiers (high, xhigh, max), function calling, structured outputs, image and PDF input, and built-in web search. On paper that reads like a normal frontier-model spec sheet. The part that isn't normal is what's happening underneath it.

RelatedKimi K3 Is the Largest Open Model Ever, and It Is Not Cheap

How is Fugu Ultra v2 different from a normal LLM release?

Most models you call through an API are a single set of weights answering every prompt end to end. Fugu Ultra v2.0 is built differently: it's a routing model trained to break a task apart and hand pieces of it to whichever model in its pool is best suited, open-weight general models, specialized models (Sakana points to coding as a priority), and additional calls back into Fugu itself. That last part, recursive self-calling, means the system can spin up further instances of its own routing logic mid-task, effectively letting an agent delegate to sub-agents that are also Fugu. Sakana says the pool deliberately excludes any individual proprietary frontier model, which is a real constraint, not a footnote. It means Fugu Ultra v2's ceiling is set by how well an orchestration layer can combine smaller and open components, not by access to GPT- or Gemini-class weights.

Why skip owning a frontier model at all?

Because Sakana's bet isn't new for them. The company has spent years on evolutionary model merging, combining existing models into new capabilities rather than training giant ones from scratch. Fugu Ultra v2.0 reads as that research applied at the orchestration layer instead of the weight-merging layer: route intelligently across many smaller specialists and let coordination do the work that a single 2-trillion-parameter model would otherwise have to do alone. If it works, the pitch is obvious. You don't need to out-train OpenAI, Anthropic, or Google on scale if you can out-coordinate a fixed-size model with layered, swappable components that get cheaper and better independently over time.

Who does this actually affect?

Developers building agentic pipelines are the most direct audience: Sakana is explicitly targeting complex multi-step reasoning, autonomous research, and full-stack software development, exactly the workloads where a single long-context call tends to fall apart. Teams currently stitching together their own AutoGPT-style orchestration on top of a single base model now have a packaged, API-level alternative that handles routing internally. It's also a data point for the frontier labs themselves: if an orchestration-first approach can compete on agentic tasks without owning a top-tier base model, that's a cheaper path to relevance than the ever-larger-monolith strategy OpenAI, Anthropic, and Google have mostly pursued.

Can we trust the "top score on 5 of 8 benchmarks" claim?

Not yet, and Sakana hasn't claimed otherwise. The benchmark results are self-reported: Sakana's own evaluation, on Sakana's own agentic benchmark suite, with no third party confirming the numbers. That's a meaningfully different claim than an independently verified score. It's worth being explicit about that gap rather than repeating the headline number as settled fact.

What happens next?

The obvious open question is whether Fugu Ultra v2.0 gets tested somewhere neutral. Coding ability is the easiest place to check: an independent SWE-bench Verified run would tell developers whether the orchestration approach holds up outside Sakana's own test suite. Latency is the other thing to watch. Routing across multiple models and recursive self-calls adds coordination overhead that a single forward pass doesn't have, and whether that overhead is worth the cost and capability tradeoff will only show up once people are running it in production.

RelatedMeta's Muse Spark 1.1 Chases Anthropic and OpenAI

Our take

The idea is more interesting than most launch-day framing suggests. Sakana isn't claiming to have built a bigger brain, it's claiming that a trained coordination layer over smaller, swappable, mostly open components can match what a monolithic frontier model does on agentic tasks, for a fraction of the commitment to owning frontier-scale weights. That's a coherent extension of the model-merging research Sakana has been doing for years, and it's a genuinely different bet than what OpenAI, Anthropic, and Google are running.

The catch is that "top or joint-top on five of eight benchmarks" is Sakana grading its own homework. We track this exact distinction on our own AI Coding Leaderboard, which scores models on SWE-bench Verified via vals.ai, an independent evaluator, specifically because vendor-reported numbers and independently verified ones tell different stories often enough to matter. Fugu Ultra v2.0 won't appear on that leaderboard until, or unless, an independent SWE-bench Verified score exists for it. Until then, the $5/$30 pricing and 1M context window are real and checkable. The performance claims are not, at least not by us.

Original analysis by GenZTech Team.