Sakana AI put two more Fugu models into service on Thursday. Max and Ultra v2 share the same orchestration engine and part ways on the objective each was tuned to hit.
Neither ships as weights. Both sit behind Sakana’s OpenAI-compatible API, and the company does not offer them in the EU or EEA. The economic argument is simple: sending a routine lookup to a trillion-parameter model wastes money, so Fugu reads a request and assembles an agent scaffold around the cheapest machinery that can still finish the job.
Max casts a wider net when it picks the models that will handle a job. Open-weight and specialized systems join the pool, NVIDIA’s Nemotron family among them under a partnership between the two companies. Sakana says the result lands close to elite models at 2x to 6x less cost, a number drawn from SWEFish, its own internal coding benchmark. Treat it as a vendor claim.
Ultra v2 goes after reasoning that runs long, autonomous research projects and full-stack software builds, and it improves most on visual and structured inputs. August 28 is its training cutoff. Anthropic’s Fable 5 and Fable 5.1 and OpenAI’s GPT-6 Astra were kept out of its agent pool by design, which Sakana says blunts exposure to vendor lock-in, revoked API access and service cutoffs.
The release rests on two papers Sakana took to ICLR 2026. One of them, TRINITY, leans on a coordinator that evolved to hand out different roles between turns, whether Thinker, Worker or Verifier. A second system, the Conductor, picks up coordination strategies written in plain language through reinforcement learning. Ship dates have come quickly: April brought the beta, June general availability, and a Claude Code interface followed in July.
