Sakana AI moved its Fugu orchestration system out of beta and into general availability on June 22, the Tokyo lab said in its launch announcement. Two months of closed testing with close to 500 users preceded the public release. The pitch is blunt: frontier-level results without depending on any single vendor.
What Fugu actually is
Strip away the framing and Fugu is a small language model trained to call other language models. Rather than a human hand-wiring roles, workflows, and which model handles what, Sakana trained a coordinator that decides at runtime who to call, in what role, with which subtasks. It can even call itself recursively, which is how the system scales effort on harder problems.
The work sits on two ICLR 2026 papers, Trinity and Conductor. Trinity is the evolved coordinator piece; Conductor handles learned natural-language coordination. If you want the mechanics rather than the marketing, those are where to look. There's also a technical report on GitHub.
You talk to all of this through one OpenAI-compatible endpoint. The complexity stays out of your code, which is the genuinely useful part.
About those benchmarks
Here's where the original buzz needs a trim. Sakana's own language is that Fugu Ultra "stands shoulder-to-shoulder" with Anthropic's Fable 5 and Mythos Preview, not that it beats them outright. And every score that isn't Fugu's own is provider-reported, with no independent evals yet. So treat the table as Sakana grading its own homework.
The wins also aren't a clean sweep. Independent benchmark coverage notes Fable 5 still tops SWE-Bench Pro and Humanity's Last Exam, GPT-5.5 leads on long-context recall, and Opus 4.8 edges the field on the CTI-REALM security test. One genuinely odd detail: on a few benchmarks the balanced Fugu scores higher than Fugu Ultra, so piling on more orchestration doesn't reliably help.
"Where other tools flag about three issues, Fugu surfaced more than twenty." That's an anonymous software engineer in Sakana's own writeup, on code review, and it's exactly the kind of testimonial a company picks because it sounds great. Useful as a signal, not as proof.
The sovereignty angle
The more interesting argument is political. Sakana leans hard on the idea that depending on one provider for critical infrastructure is a material vulnerability, pointing at the recent export controls on Fable and Mythos as proof access can vanish overnight. Because Fugu orchestrates swappable agents, the claim goes, it routes around any single provider getting cut off. Neither Fable 5 nor Mythos Preview is even in Fugu's pool, since they aren't publicly accessible.
It's a clean story. It also conveniently sells a product. Critics have pointed out the system relies on undisclosed mixes of closed frontier models without saying what share is open versus closed, which undercuts the sovereignty framing somewhat.
What it costs
Sakana confirms two access paths: subscription tiers for everyday use and pay-as-you-go for heavier and enterprise workloads, both Fugu and Fugu Ultra behind the same API. Specific per-token rates aren't published on the release page, so anyone quoting exact dollar figures should check the console before budgeting around them.
Fugu and Fugu Ultra are live now through the console site. The EU and EEA rollout reportedly lags the rest, so European teams should confirm availability before integrating.
