Sakana AI released Fugu Max and Fugu Ultra v2 on 11 September, reporting 48.3 on Chartography against 27.3 for Opus 5, and Fugu Max at $2/$6 per million tokens. All figures are Sakana's own. Fugu is an orchestrator, not a model: it routes across a pool of open and specialised models and recursively calls itself, and every hop is metered. That makes the posted per-token rate an answer to the wrong question, and the gap widens on exactly the hard tasks where the benchmark wins are. Sakana's supply-chain argument survives the critique.
Questions
An orchestrator is a system that decomposes your task, routes the parts to different underlying models chosen for each part, and assembles the results. Sakana's Fugu additionally can call instances of itself recursively. Calling a cheaper model is one decision made once; orchestration is a sequence of decisions made per task, each of which is itself an inference call that consumes billed tokens.
Go further
We read the release notes and the rate cards together, including the parts vendors put in captions.
Read AI coverage
DrafterDaily Editorial
Keep reading
An evaluation company published a decoding of a 17th-century cryptogram. A refutation with a hash manifest and a re-runnable script landed the next day. The claim spread two weeks later; the refutation did not.
Two Colorado AI statutes take effect on 1 January 2027. One obliges the Attorney General to write implementing rules. The other does not — and is getting them regardless. That asymmetry decides where the compliance work lives and where a legal challenge would aim.
Anthropic's September threat report documents two automation loops in enough detail to reason about. Both have a chokepoint the speed coverage misses — and the 395 figure circulating in secondary reports is not in the primary source.