Sakana Just Sold the Orchestrator as a Model. It Beat Opus 5 on Chartography at Two Dollars a Million.

Sakana AI shipped Fugu Max v1.0 and Fugu Ultra v2.0 on Thursday, September 11, 2026. The launch post is short and the sentence that decides the read is one line: Fugu is a language model whose only job is to route tasks across a pool of open-weight and specialist sub-models, and to call instances of itself recursively when the plan needs more compute. Max lists at $2 per million input and $6 per million output tokens. Ultra v2 lists at $5 and $30. On Chartography, a visual reasoning benchmark, Fugu Ultra v2 scores 48.3 against Opus 5 at 27.3 and Fable 5 at 29.5. That is a Japanese lab, no frontier weights of its own, pricing an orchestrator below every frontier line on the pricing page and beating two of them on a public benchmark at the same time.

Two weeks ago the harness thesis at TensorFeed was still a forward read on a curve. Nine days ago Anthropic shipped the Fermat receipt on Prove2Me. Two days ago OpenAI put the Codex harness behind one API call and priced the orchestrator at zero. Yesterday morning Sakana priced the orchestrator at two dollars a million and wrapped it around models it does not own. Three receipts in nine days, three different pricing postures, three different theories of where the margin lives. The Sakana one is the odd receipt out, and it is the one that will get read hardest by anyone building a coding product on someone else's API tier.

The Pricing Line

Line up the top of the agent-tier pricing page as of today.

The Fugu Max output line is $6. The nearest frontier output line is Fable 5.1 at $50, an eight times spread. The Ultra v2 output line at $30 is 60 percent of Astra output on a model that beats it on Chartography and ties or wins DeepSWE against models three to five times its price on Sakana's own reported figures. Neither Fugu tier is priced against the frontier. Both are priced against the mid-tier line: Sonnet 5, Gemini Flash, Kimi K3. The read is not that Sakana is cheaper than frontier, it is that Sakana has decided the orchestration layer belongs on the mid-tier price shelf and the sub-models it dispatches to are cheap enough to fit inside that number.

What Is Actually Inside Fugu

Two ICLR 2026 papers do the work. TRINITY is a coordinator on the order of 0.6 billion parameters, evolved with CMA-ES, that assigns Thinker, Worker, and Verifier roles across a pool of larger models. Conductor is a 7 billion parameter model trained with reinforcement learning to discover natural-language coordination strategies, and it can call itself recursively to scale test-time compute. Fugu Max and Fugu Ultra v2 productize both. The orchestrator is small. The sub-models it calls are not.

The pool includes NVIDIA's Nemotron family through a named collaboration, an unspecified set of open-weight coders (Qwen and DeepSeek shapes on the public benchmark traces), and Sakana's own specialist models. What Fugu does not call, on the day one card, is any proprietary frontier model. That is a design choice with a margin behind it. Every token routed to Nemotron or Qwen is a token Sakana bills the customer for and pays an open-weight inference vendor for at commodity rates. Every token routed to a proprietary API would be a token Sakana bills for and pays OpenAI or Anthropic list price for. The math only works with open weights underneath, and the pricing sheet is the public form of that math.

The Benchmark Sheet

Fugu Max is the best overall score on six benchmarks: Terminal Bench 2.1, GPQAD, AA-LCR, GDP.pdf, AutomationBench, and SWEFish. Fugu Ultra v2 tops or ties five of eight: GDP.pdf, Chartography, DeepSWE, Toolathon, and SWEFish. Chartography is the interesting row because it is a visual reasoning benchmark and the sub-model doing the vision work is not a Sakana model, it is whatever the orchestrator dispatches to on that call. Ultra v2 scores 48.3, Opus 5 scores 27.3, Fable 5 scores 29.5. On a benchmark where the frontier labs have a native vision stack and Sakana does not have a vision model at all, the orchestrator wins by roughly 20 points against the pair Anthropic ships. Either the benchmark is soft, or the orchestrator is doing something the frontier vision stacks are not, and the papers say it is doing something specific (splitting the image into a chain of role-assigned queries against sub-models and reconciling the outputs against a verifier).

Two caveats attached. SWEFish is Sakana's own internal benchmark and its top-scoring rows should be read as a self-reported comparison until an independent evaluator posts a number. The public rows (Terminal Bench 2.1, GPQAD, DeepSWE, Chartography) have run against Anthropic and OpenAI submissions in the same period, and those are the ones that set the read. Even inside the public rows, orchestrator wins are not a claim about the underlying weights, they are a claim about the coordination policy the orchestrator has learned.

The Harness Thesis, Third Receipt

Line up the three receipts in a table.

Four rows, four pricing postures. Anthropic and OpenAI put the harness on the research site to seed the market and then either kept it (Prove2Me still lives inside Anthropic Research) or shipped it under the token line (OpenAI Agents API). Sakana skipped the research-post stage and went straight to a priced SKU on OpenRouter and its own console. The Sakana version is the honest form of the harness thesis. If the value in a long-horizon coding workload is the coordination policy rather than the weights, then the price of the workload should be the price of the policy plus the wholesale cost of the weights. Fugu Max at $2 and $6 is the first line item on any pricing page that reflects that decomposition.

What This Does to Frontier Margin

Less than the headlines will say, more than the frontier pricing pages currently reflect. Sakana is one lab, its Nemotron partnership is one collaboration, and the sub-model pool today is heavy on open weights that already run below list price on any half-serious inference vendor. What is new is that a paying customer now has a public API to compare against on coding, terminal, and visual reasoning workloads, and the API costs a quarter to a tenth of the frontier line on output tokens. Every enterprise procurement team that runs a bake-off in Q4 has a new column to fill in.

The pricing floor read from our pricing war piece moves a step. The floor was set by frontier cache-read cuts and mid-tier commoditization. The Sakana receipt adds a third pressure: an orchestrator can charge a premium over commodity inference while still undercutting frontier on the total invoice, because the orchestrator is small and the sub-models are cheap. That is a different curve from the one Fable 5.1 and Astra sit on, and it fits inside the same coding-agent budget without asking the buyer to swap the whole stack.

Two structural effects worth pricing. First, if Fugu-shape products proliferate (Sakana plus a second lab plus a US independent) the frontier labs lose the argument that their proprietary weights are the only credible substrate for a coding agent, because the public counter-example runs on Nemotron and Qwen. Second, the sub-model market gets new demand. Nemotron just picked up a routed workload from a Japanese lab with a paying API, and that is a data point for every open-weight vendor pitching hyperscalers on being the default inference target for the next Fugu-shape entrant.

What It Does Not Do

Three caveats worth naming because the take will run hot for a week. First, orchestrator wins on benchmarks are not the same as orchestrator wins on production workloads. A real coding agent needs long-horizon memory, tool integration, sandbox access, and error recovery across sessions, and Fugu at launch is a hosted model API not an IDE. Second, the sub-model pool is a supply-chain dependency Sakana does not fully control. Nemotron licensing is generous today, Qwen is open weights today, both could tighten. Third, the recursive self-call architecture increases token consumption in ways the launch post does not fully disclose. A Fugu Max call that recurses three times against a Nemotron backend is not one $6 output line, it is several, and the effective price against a monolithic frontier call is a workload-shape question that has to be measured not read off the page.

Our Take

The interesting decision is where Sakana put the money. Anthropic monetized the harness by keeping it inside its own research group and letting the Claude token line carry the margin. OpenAI monetized the harness by pricing it at zero and letting Astra carry the margin. Sakana monetized the harness by making the harness itself the paid product and letting the sub-models settle at commodity rates. All three are defensible, only one survives contact with a customer who wants a coding agent on Nemotron because Nemotron already runs in their private cloud. If the next two Fugu-shape products come from a US lab and a European lab in the same quarter, the frontier margin story on coding workloads has a hole in it that a token repricing does not fill.

Practical read for builders on the API: if your product hits a wall on latency, cost, or the coordination policy the frontier tier gives you out of the box, Fugu Max is a real bake-off candidate at a quarter of the output price, and the sub-model pool means you are not betting the workload on one weights owner. The counterpoint is that Fugu at launch is a hosted service without the ecosystem of tools, sandboxes, and IDE integration OpenAI shipped alongside Agents API on Sep 10, so a coding-agent product shopping today has to weigh a cheaper token line against a thinner integration surface, and that trade-off resolves differently at different points of the stack.

Three Signposts for the Next 60 Days

Whether a US or European lab publishes a Fugu-shape orchestrator with a public price within the next 60 days (the direct test of whether the orchestrator-as-a-model pattern generalizes past Sakana, or stays a Japanese one-off tied to the Nemotron partnership). Whether Anthropic or OpenAI publishes a coordination-policy paper of their own with a benchmark row alongside the frontier model number (the direct test of whether the frontier labs concede that the harness is a separately priceable asset, or absorb the claim inside the token line). Whether an independent evaluator (Vellum, Artificial Analysis, an academic group) posts a public re-run of the DeepSWE and Chartography rows against the same models in the same window (the direct test of whether Sakana's self-reported benchmark sheet holds up under third-party replication). Two of the three fire and the harness thesis moves from a TensorFeed read into a category on the pricing page, and the frontier margin story on coding workloads gets rewritten inside a quarter.

Marcus Chen, September 13, 2026. TensorFeed tracks AI model releases, provider status, and the pricing layer underneath them. Sources for this piece: Sakana AI's September 11 launch post for Fugu Max and Fugu Ultra v2, the TRINITY and Conductor ICLR 2026 papers, VentureBeat and MarkTechPost coverage of September 10 and 11, and the Anthropic Fermat and OpenAI Agents API announcements referenced above.