Sakana AI turns multi-agent swarms into a single model API

Relying on a single AI provider got a lot scarier the day export controls knocked Anthropic's top-tier models off the table for global users. When rules shift overnight, your entire production stack can evaporate with them.

Tokyo-based Sakana AI thinks it has the antidote. The lab launched Fugu and Fugu Ultra, pitching a multi-agent orchestration setup that behaves like a lone foundation model behind a single OpenAI-compatible API.

One endpoint to rule them all

Instead of building another giant monolithic model from scratch, Sakana trained Fugu to act as a coordinator. It evaluates your prompt, decides whether to answer directly, or delegates pieces of the problem to a pool of swappable expert models.

All the delegation, cross-checking, and final synthesis happen behind the curtain. Your code only ever talks to one endpoint.

At launch, users get two flavors. Standard Fugu is tuned for low-latency coding and day-to-day chat, and even lets privacy-conscious teams opt specific models out of the pool. Fugu Ultra is built for heavier lifting, like scientific research, patent reviews, and automated cybersecurity audits. An update also added Fugu Max to balance the cost-performance Pareto frontier.

Benchmark highs and real-world skepticism

On paper, the numbers are wild. Sakana claims Fugu Ultra matches or nears Anthropic's Fable 5 and Mythos Preview across engineering and reasoning benchmarks, while outperforming GPT 5.5, Gemini 3.1 Pro, and Opus 4.8 on complex multi-step tests.

Beta testers highlighted impressive endurance. One cybersecurity tester reported end-to-end vulnerability scanning and reporting without drifting out of scope, while a software engineer praised it for flagging over twenty code issues where typical tools found three.

Yet early reception outside the lab has been decidedly mixed. Critics pointed out the irony that Fugu delivers frontier performance primarily by leaning on other frontier models. Meanwhile, users have noted that real-world output does not always match the scoreboard, raising questions about pricing and the opaque model mix.

Why it matters

We think Sakana is smartly capitalizing on enterprise paranoia over vendor lock-in and geopolitical export controls. Positioning orchestration as a hedge against regulatory bans is clever marketing that taps into genuine enterprise dread.

For researchers and security analysts tackling messy, multi-step workflows, Fugu Ultra is worth testing on a pay-as-you-go basis just to skip the headaches of writing custom multi-agent routing from scratch.

However, we believe the lack of visibility into what is running under the hood remains a massive hurdle. Paying a premium for an orchestration layer without knowing which underlying model did the work, or how many tokens were burned, makes it a hard sell for cost-sensitive builders.