Sakana AI Ships Fugu Max and Fugu Ultra v2, Orchestrator Models Sold Through a Hosted API

Fugu is not a single foundation model. It is a learned orchestrator that routes work across a pool of other models behind one API. The new release keeps that architecture and splits it along two optimization targets: Fugu Max aims at the best output per dollar, while Fugu Ultra v2 aims at the highest capability on hard, multi-step tasks. Sakana argues that real workloads are judged on capability and cost together, and describes the trade-off as a Pareto frontier on which higher quality costs more and lower cost costs quality.

The release follows a fast cadence for the family. Fugu entered beta in April, reached general availability in June, and added Fugu-Cyber and a Claude Code interface in July. Sakana's technical report describes Fugu models as language models in their own right that read a query and build an agentic scaffold for it on the fly. Training combines large-scale fine-tuning, evolutionary algorithms and reinforcement learning, and the system builds on two ICLR 2026 papers: TRINITY, which uses a lightweight evolved coordinator to assign Thinker, Worker or Verifier roles across turns, and The Conductor, trained with reinforcement learning to discover natural-language coordination strategies and focused prompts.

Fugu Max widens the pool of models Fugu can orchestrate, adding a large set of open-weights and specialized models, including the NVIDIA Nemotron family through a collaboration with NVIDIA. Each task is routed to the leanest model capable of solving it. Sakana prices Fugu Max at $2 per million input tokens and $6 per million output tokens, which it says is 40% to 60% lower on output price than Sonnet 5, GPT 5.6 Terra and Kimi K3. The company reports the best overall score on six benchmarks — Terminal Bench 2.1, GPQA Diamond, AA-LCR, GDP.pdf, AutomationBench and SWEFish — and says the model expands the cost-performance Pareto frontier on seven of ten benchmarks, placing it within striking distance of elite models at two to six times lower cost. SWEFish is an internal Sakana benchmark built from the company's own coding challenges.

Fugu Ultra v2 targets complex reasoning, autonomous research and full-stack software development, with its largest gains reported on sustained reasoning over visual and structured data. It scores 48.3 on Chartography, a visual reasoning and data interpretation benchmark, against 27.3 for Opus 5 and 29.5 for Fable 5, and 74.3 on DeepSWE, a real-world software engineering benchmark, ahead of models priced three to five times higher per token. Sakana reports best or joint-best results on five of eight benchmarks — GDP.pdf, Chartography, SWEFish, DeepSWE and Toolathon — and top-two placement on seven of eight. Fable 5, Fable 5.1 and GPT-6-Astra are not in Fugu Ultra v2's agent pool. The model's training cutoff is August 28, 2026.

Sakana's stated aim is frontier-level output without dependence on any single proprietary model, which the research team says reduces exposure to vendor lock-in, API revocations and sudden service cutoffs. Existing users can switch to either model with a one-line change to their API configuration.

Editor's Summary

Sakana AI's Fugu Max and Fugu Ultra v2 extend its orchestrator approach, in which one API dispatches tasks to a pool of third-party models rather than relying on a single foundation model. Fugu Max competes on price and breadth of routing options, while Ultra v2 targets the hardest reasoning and software engineering benchmarks, in both cases without the proprietary models Sakana names as absent from its pool. The models ship only as a hosted API, with no open weights and no availability in the EU or EEA.