11 SEP 2026 — Sakana AI released two models on Thursday, and both are orchestrators: models trained to route a task across a pool of other models rather than answer it alone. The interesting part is what one of those pools leaves out. Fugu Ultra v2 excludes Fable 5, Fable 5.1 and GPT-6-Astra, and still posts 48.3 on the Chartography visual-reasoning benchmark, against 27.3 for Opus 5 and 29.5 for Fable 5.
The second release, Fugu Max, is the cheap one: two dollars per million input tokens and six per million output, which Sakana puts at 40 to 60 per cent below Sonnet 5, GPT 5.6 Terra and Kimi K3.
What Fugu is, and is not
Fugu is not a frontier model, and it does not compete on parameter count. It is an orchestrator trained to route work to other models. It reads a query, builds an agentic scaffold for that query on the fly, routes sub-tasks across a fixed pool of open-weight and specialised models, and can recursively call instances of itself.
Two Sakana research lines sit underneath it. TRINITY evolved a lightweight coordinator that assigns Thinker, Worker and Verifier roles across turns. The Conductor was trained with reinforcement learning to discover coordination strategies in natural language rather than having them specified. The shipped system combines large-scale fine-tuning, evolutionary search and reinforcement learning.
A Fugu score, then, is a score for a system rather than for a set of weights. The distinction is critical to reading the performance numbers below, and the release does not foreground it.
The exclusion is the product claim
Sakana states plainly that Fugu Ultra v2 reaches its scores without Fable 5, Fable 5.1 or GPT-6-Astra in its agent pool, and frames the omission as reduced exposure to vendor lock-in, API revocations and sudden service cutoffs.
This is a supply-chain argument, not a research result. An orchestrator whose best answers require a call to one vendor's flagship is that vendor's reseller, with the vendor's pricing, availability and terms passed straight through to the customer. An orchestrator that scores well on a pool of open-weight and specialised models has a different risk profile, and the risk profile is the part a buyer can actually act on.
A token price is not a query price
The release does not publish how many model calls a Fugu answer takes, or how long it waits for them.
A per-token price and a query cost are close to the same thing for a single model. For an orchestrator they are not. A system that builds a scaffold, routes sub-tasks to several models and can recursively call itself is by construction spending more calls per query than a single model does. Whether those calls are inside the quoted price or alongside it is the difference between a 40 to 60 per cent saving and a headline.
Sakana's framing suggests the former — it describes routing each task to the leanest model capable of solving it, which is a cost-reduction mechanism rather than a cost-shifting one, and it claims to expand the cost-performance Pareto frontier on seven of ten benchmarks, which is a claim about cost and score together. That is the right shape of claim. It is also not the same as publishing latency and call counts, and neither figure appears.
Anyone evaluating this should measure their own workload rather than reading the per-token line. The orchestration premium falls on queries that fan out, which are exactly the queries an orchestrator is being bought for.
What the benchmark table says
Fugu Max claims best overall scores on six benchmarks: Terminal Bench 2.1, GPQA Diamond, AA-LCR, GDP.pdf, AutomationBench and SWEFish. Fugu Ultra v2 claims best or joint-best on five of eight and a top-two placing on seven of eight, with DeepSWE at 74.3.
Two cautions apply to any release of this kind, not only Sakana's. The benchmark set is chosen by the party announcing the result, and "best on six" is a statement about six benchmarks someone selected. And the Chartography gap — 48.3 against 27.3 — is large enough to be worth treating as a question rather than a conclusion. A 21-point margin over a frontier model on a visual-reasoning task suggests either a capability difference or a difference in method. A system that can route a chart to a specialised vision model is, by design, using a different method.
This is not a criticism; routing to the right specialist is the entire thesis. The comparison, however, is between a system and a model.
No weights, no Europe
Two constraints matter more than the scores for most readers.
There are no open weights. Fugu is available only as a hosted API behind an OpenAI-compatible interface, so the vendor-lock-in argument that justifies the agent pool's composition does not extend to Fugu itself. A buyer swaps dependence on one frontier vendor for dependence on Sakana.
And the service is not available in the EU or EEA. For readers in this region that is a quiet advantage: an ASEAN team can evaluate Fugu on the day it ships while a European competitor cannot, which is an inversion of the usual order and worth noticing while it lasts. The training cutoff is 28 August 2026.
What to watch
The claim to test is not whether Fugu beats Fable 5 on a chart, but whether an orchestrated pool of smaller models holds up on the long, ambiguous agentic work where the routing decision is itself hard. Benchmarks reward systems that can recognise the task type. Production does not always announce it.
The second thing to watch is the pool. Sakana's collaboration with NVIDIA puts the Nemotron family into Fugu Max, and a pool that depends on which open-weight models exist is a pool that changes when the open-weight frontier moves. As long as labs keep publishing weights, that is a strength. If they stop, it becomes a dependency of a different kind.
View author profile → · Editorial policy
