Confirmed in a primary source 11 Sept 2026 Models

A Japanese company released a system it says beats the most expensive models without using any of them.

On Friday Sakana AI released two versions of Fugu, a product that does not answer questions itself but routes each task to whichever model can do it best for the money. Fugu Max is priced at 2 dollars per million input tokens and 6 dollars per million output tokens, which the company says is 40 to 60 percent below Anthropic's Sonnet 5, OpenAI's GPT 5.6 Terra and Moonshot's Kimi K3, and the larger Fugu Ultra v2 claims the best or joint-best score on five of eight benchmarks while excluding Fable 5, Fable 5.1 and GPT-6 Astra from the models it is allowed to call.

The caveat is that every number here is the company's own evaluation published on its own site, and no independent run of any of it exists.

Until now

The frontier model was the product, and anything built on top of it was a wrapper that inherited its price.

Now

Put more weight on routing being a product that can beat the models it routes to, and on frontier pricing power resting on being irreplaceable inside somebody else's system; put less weight on benchmark leadership belonging to whoever trained the biggest model.

Sakana AI's product does not know anything. It decides. Given a task, Fugu picks which model should handle it and pays for that one, and the company's argument is that choosing well beats being the best single model. On Friday it shipped two versions. Fugu Max costs 2 dollars per million tokens in and 6 out, which Sakana puts at 40 to 60 percent below Anthropic's Sonnet 5, OpenAI's GPT 5.6 Terra and Moonshot's Kimi K3, and it claims the top score on six named benchmarks. Fugu Ultra v2 is claimed at 48.3 on Chartography, where Anthropic's Opus 5 scores 27.3 and Fable 5 scores 29.5, and at 74.3 on DeepSWE.

The detail carrying the argument is an exclusion. Sakana says Ultra v2 reaches those results without Fable 5, Fable 5.1 or GPT-6 Astra in the pool it can call, which is a claim that you can beat the three most expensive models in the world without buying any of them. A day earlier Cognition made the same price argument another way, shipping its coding model post-trained on a Chinese open-weight base at 64 percent below Anthropic. Both are company numbers on company benchmarks and neither has been run by anyone else. Put more weight on routing being a product that can beat the models it routes to, and on frontier pricing power resting on being irreplaceable inside somebody else's system, and less on benchmark leadership belonging to whoever trained the biggest model.

How it got here

Jul 27
Moonshot AI publishes Kimi K3 with open weights, which makes a capable model free to route work to.
Sep 10
Cognition ships SWE-2, post-trained on Kimi K3, within one point of Anthropic's Fable 5.1 at 64 percent lower cost. Carded in Edition 018.
Sep 10
DeepSeek releases V4.1-Flash with native vision and cuts its API prices, per its own changelog.
Sep 11
Sakana AI releases Fugu Max at 2 and 6 dollars per million tokens and Fugu Ultra v2, claiming best or joint-best on five of eight benchmarks without Fable 5, Fable 5.1 or GPT-6 Astra in the pool.
Who has to react
Anthropic and OpenAI, whose most expensive models are named as the ones not needed; Nvidia, whose Nemotron family sits in the pool Sakana routes to and which sells the hardware either way; Moonshot AI, whose open weights make this economics possible and which Anthropic accuses of copying Claude to obtain them; the maintainers of Terminal Bench, Chartography and DeepSWE, now cited in a pricing argument; enterprise buyers who signed annual commitments at frontier list prices; and any independent evaluator willing to run these numbers, because nobody has.

The evidence

ObservedSakana AI published the release on its own site on Friday September 11, 2026, Japan time. We read the post.

The whole observation, as recorded

Two models: Fugu Max, tuned for cost and performance, and Fugu Ultra v2, tuned for capability on hard tasks. Both are orchestrators, meaning the product decides which underlying model handles a task rather than answering itself. Fugu Max is priced at 2 dollars per million input tokens and 6 dollars per million output tokens, which the company states is 40 to 60 percent below the output pricing of Anthropic's Sonnet 5, OpenAI's GPT 5.6 Terra and Moonshot's Kimi K3. Sakana claims Fugu Max takes the best score on Terminal Bench 2.1, GPQAD, AA-LCR, GDP.pdf, AutomationBench and SWEFish, and that it expands the cost-performance frontier on seven of ten benchmarks; its pool includes open-weight and specialised models, among them Nvidia's Nemotron family. Fugu Ultra v2 is claimed at 48.3 on Chartography against 27.3 for Opus 5 and 29.5 for Fable 5, and at 74.3 on DeepSWE, best or joint-best on five of eight benchmarks and in the top two on seven of eight, with a training cutoff of August 28, 2026. Sakana states that Ultra v2 reaches these results without Fable 5, Fable 5.1 or GPT-6 Astra in its agent pool. Both models are available immediately through an OpenAI-compatible interface, which the company says existing users reach with a single parameter change. Its stated claim is that orchestration consistently outperforms isolated models. Every figure on this card is Sakana's own and we found no independent evaluation as of this morning. Sakana AI was founded in Tokyo in 2023 by former Google researchers, including a co-author of the transformer paper. Edition 018 carded Cognition's SWE-2, which came within one point of Anthropic's Fable 5.1 at 64 percent lower cost by post-training on Moonshot's Kimi K3; this is the same price argument reached by a different method. American coverage of this release is dated September 10 and Sakana's own post is dated September 11 in Japan; we date it by the company.

Established

Sakana AI released Fugu Max and Fugu Ultra v2 on September 11, 2026, available immediately through an OpenAI-compatible interface.
Fugu Max is priced at 2 dollars per million input tokens and 6 dollars per million output tokens.
Sakana states Fugu Ultra v2 scores 48.3 on Chartography against 27.3 for Opus 5 and 29.5 for Fable 5, and 74.3 on DeepSWE.
Sakana states Ultra v2 achieves its results without Fable 5, Fable 5.1 or GPT-6 Astra in its agent pool.

Not established

That the scores hold. Every figure is the vendor's own and no independent run exists.
That it is cheaper in use. A router that calls several models per task can cost more than its headline rate.
That frontier models are replaceable generally. Eight benchmarks are not the set of things people buy models for.
That Sakana makes money at this price. No revenue, volume or margin figure was published.

Why it weighs this much

Signal strength

CostMedium

CommitmentHigh

ReversibilityModerate

PersistenceMonth

Building and shipping two models costs real money, and the published price is the part that commits: a company that names 2 and 6 dollars per million tokens cannot quietly charge more next week. Cost is medium rather than high because an orchestrator is far cheaper to build than a frontier model, which is the whole point of building one. Reversibility is moderate because prices can be raised and models withdrawn, and this industry has done both this year. Persistence is a month, deliberately low, because every number on this card is Sakana's own and a single independent evaluation could remove the reason to care. The bars stay here until somebody outside the company runs these benchmarks.

What to watch

strengthens An independent evaluator or third-party leaderboard reproduces Fugu Ultra v2 at 45 or above on Chartography, or above 70 on DeepSWE, by March 31, 2027.
weakens Sakana raises Fugu Max list pricing above 3 dollars per million input tokens, or withdraws either model, by June 30, 2027.

The words

Orchestrator: Software that takes a task and picks which model does it, instead of answering itself. What it sells is the choice, not the knowledge.
Token: The unit models are billed in, roughly three quarters of a word. Prices are quoted per million tokens sent in and per million produced.
Open weights: A model whose parameters anyone can download and run. It is what makes a cheap pool of models to route to possible at all.
Benchmark: A fixed set of test tasks used to score models. Vendors run their own and publish them, which is why an independent run is the thing that settles it.

Your call

Which explanation do you believe?

A call is final. Other people's calls show after yours.