Sakana Fugu Ultra v2: A Model That Runs Other Models
Sakana AI's Fugu Ultra v2 and Fugu Max orchestrate other models instead of replacing them. What the split means for agent cost, quality, and lock-in.
Sakana AI released Fugu Max and Fugu Ultra v2 on 10 September 2026 — two versions of a single orchestrator, a system that answers a request by hiring other models to do the work. The split is simple: Fugu Max chases the best output per dollar, Fugu Ultra v2 chases the highest capability, and neither needs a single frontier model in its pool to get there.
That makes it more than one lab's release notes. Orchestration is the layer your personal agent actually lives in — it is the part that decides which model burns your money on a task a cheaper one could have finished.
What shipped
- Both models went live on 10 September 2026 through Sakana's OpenAI-compatible API, and existing Fugu users switch with a one-line parameter change, according to Sakana's release post.
- Fugu Max is priced at $2 per million input tokens and $6 per million output tokens.
- Fugu Ultra v2 runs $5 input and $30 output per million, with $0.50 cached input, and higher rates above a 272K context window.
- There are no open weights to self-host, and Sakana does not offer the service in the EU/EEA, as MarkTechPost's write-up notes.
The pricing line is where this stops being abstract. Fugu Max's output price is reported 40–60% below Sonnet 5, GPT 5.6 Terra and Kimi K3.
The idea: a model that is really a team
Fugu is not a foundation model with a bigger brain. It is a learned orchestrator: it reads a request, builds an agentic scaffold for that specific task, and routes each step to the model best suited to it. Sakana trains it with a mix of large-scale fine-tuning, evolutionary methods and reinforcement learning, building on two ICLR 2026 papers — one where an evolved coordinator assigns Thinker, Worker and Verifier roles across turns, and one where a conductor learns coordination strategies in natural language.
The practical consequence is that the model pool is swappable. Fugu Max deliberately widened its pool to include more open-weights and specialised models, including NVIDIA's Nemotron family through Sakana's collaboration with NVIDIA. That is the difference between an architecture and a product: if one model in the pool disappears, the orchestration continues.
Where each model is meant to be pointed
| Fugu Max | Fugu Ultra v2 | |
|---|---|---|
| Optimised for | Best output per dollar | Highest capability |
| Published output price | $6 per 1M tokens | $30 per 1M tokens |
| Benchmark claim | Best overall on six benchmarks | Best or joint-best on five of eight |
| Model pool | Largest to date, including NVIDIA Nemotron | No Fable 5, Fable 5.1 or GPT-6 Astra |
| Point it at | High-volume, repeatable agent work | Multi-step reasoning, research, full-stack code |
The numbers behind those claims: Fugu Ultra v2 scores 48.3 on Chartography (visual reasoning and data interpretation) against 29.5 for Fable 5 and 27.3 for Opus 5, and 74.3 on DeepSWE for real-world software engineering — ahead of models priced three to five times higher per token, per Sakana. Fugu Max takes the best overall score on six benchmarks including Terminal Bench 2.1 and AutomationBench.
Where each Fugu model wins — benchmark gap and the token bill at 5M output tokens
What it means for your agent's bill
An agent that runs on a schedule — the morning briefing, a price watch, a mailbox triage — is a token meter running in the background whether or not the answer needed a frontier model. A one-line summary of three emails does not deserve the same model as a multi-step research task, and most agents have no mechanism for telling the two apart.
That is exactly the problem orchestration solves, and it is why model routing is one of the most consequential settings in any agent setup: it decides whether your monthly bill tracks your usefulness or your curiosity. A system like Fugu is a routing decision you delegate rather than write — and if you run your own agent, the routing and cost controls are documented at docs.wolffi.sh.
The honest caveats
- These are vendor benchmarks. SWEFish is Sakana's own internal benchmark suite, and "best overall on six benchmarks" is a claim by the company selling the tokens. Treat the direction as real and the exact margins as marketing until independent runs appear.
- Your workload is not a benchmark. A 48.3 on visual reasoning says nothing about whether the orchestrated answer to your request is better than a cheap model's. Route one real task each way and compare the output, not the score.
- Lock-in moved, it did not vanish. Sakana's stated advantage is not depending on Fable 5, Fable 5.1 or GPT-6 Astra. But Fugu itself is a hosted API with no open weights and no EU/EEA availability — you have traded one dependency for another, with a different failure mode.
- The training cutoff is 28 August 2026. Orchestration decisions about which model is "leanest" can only be as current as that.
The takeaway
Fugu Ultra v2 and Fugu Max are the clearest statement yet that the useful unit of an AI system is no longer a single model — it is the layer that chooses between them. For anyone running a personal agent, the lesson is portable even if you never call Sakana's API: decide per task how much model you need, write that rule down, and let the boring work stay cheap. The agents that survive contact with a monthly bill are the ones that were never expensive by default.
