Sakana AI shipped Fugu Max and Fugu Ultra v2.0 on September 11, 2026. Sakana Fugu Max costs $2 per million input tokens and $6 per million output — 40% to 60% below Sonnet 5, GPT-5.6 Terra and Kimi K3, according to Sakana. It is not a foundation model. It is a router that sends each piece of a task to the cheapest model that can still finish it.
That distinction is the whole story. Every price cut this year came from a lab training a cheaper model. This one comes from a company that trained a dispatcher instead.
What is Sakana Fugu Max?
Sakana Fugu Max is a learned orchestrator. A single OpenAI-compatible endpoint accepts your request, breaks it into subtasks, and routes each one to a pool of other models — open-weight models, specialized models, and NVIDIA’s Nemotron family. You never pick the worker. Fugu does.
How the orchestrator actually works
MarkTechPost reports the system is built on two ICLR 2026 papers: TRINITY, which assigns Thinker, Worker and Verifier roles, and The Conductor, a reinforcement-learning coordination layer.
Training combined large-scale fine-tuning, evolutionary algorithms and reinforcement learning. The coordinator can also call instances of itself recursively, according to Pondero’s writeup.
The practical consequence: model selection moves inside inference. Sakana can swap the worker pool without breaking a single line of your application code.
What shipped on September 11
Two tiers. Fugu Max is the cost-optimized variant with the widest worker pool. Fugu Ultra v2.0 is the quality tier with a fixed, deeper pool.
Both carry a 1 million-token context window and a 128,000-token maximum output, per DataNorth. Both are hosted API only — there are no downloadable weights — and neither is available in the EU or EEA.
Sakana’s training cutoff is August 28, 2026. Switching between the two tiers takes a single parameter change.
How much does Sakana Fugu Max cost?
Fugu Max runs $2 per million input tokens and $6 per million output, with cached input at $0.25. Fugu Ultra v2.0 runs $5 and $30. Above 272,000 tokens of context, Ultra v2 jumps to $10 input and $45 output — a surcharge that catches long-document workloads.
| Tier | Input / 1M | Output / 1M | Cached input / 1M |
|---|---|---|---|
| Fugu Max v1.0 | $2.00 | $6.00 | $0.25 |
| Fugu Ultra v2.0 | $5.00 | $30.00 | $0.50 |
| Fugu Ultra v2.0 (272K+ context) | $10.00 | $45.00 | $1.00 |
Web tool calls add $0.007 each. Subscription tiers sit at $20 Standard, $100 Pro and $200 Max.
For comparison, our breakdown of Claude Fable 5.1 pricing put frontier input at $10 per million. Fugu Max is asking one-fifth of that.
Is Fugu Ultra v2 better than Opus 5?
On Sakana’s own numbers, on some tasks, yes. Fugu Ultra v2 scored 48.3 on Chartography against Opus 5’s 27.3 and Fable 5’s 29.5 — a 77% margin over Opus 5. It posted 74.3 on DeepSWE. Sakana claims top-two placement on seven of eight benchmarks tested.
Fugu Max, the cheap tier, reportedly finished best overall on six benchmarks: Terminal-Bench 2.1, GPQA Diamond, AA-LCR, GDP.pdf, AutomationBench and SWEFish.
Sakana says it expanded the cost-performance frontier on seven of ten benchmarks.
Read that with one hand on your wallet. Every one of those figures is self-reported, and the Fugu Max comparison set was explicitly bounded to models “in a similar price range.” A router that wins against cheap models is not the same claim as a router that wins outright.
It is also worth noting what a routing score measures. Chartography at 48.3 says the dispatcher picked good workers, not that any single model got smarter. The intelligence is in the assignment.
Why did Sakana exclude Fable 5 and GPT-6-Astra?
Deliberately, and it is the most commercially loaded decision in the launch. Fugu Ultra v2 hit 48.3 on Chartography and 74.3 on DeepSWE with Fable 5, Fable 5.1 and GPT-6-Astra removed from its worker pool. Sakana frames this as reducing exposure to vendor lock-in, API revocations and sudden service cutoffs.
That is not a hypothetical risk. We covered the OpenAI Cursor cutoff, where models went dark on a fixed date for a major customer.
Sakana is selling the absence of the two biggest American labs as a feature. Buyers who watched a supplier turn off the tap will hear it.
The trade is real, though. You are not removing a dependency — you are moving it. Instead of depending on OpenAI, you depend on a Tokyo company that decides, invisibly, which model handled your request today.
Who wins and who loses financially?
The losers are the mid-tier frontier APIs. If a router at $2 per million matches Sonnet 5, GPT-5.6 Terra and Kimi K3 on the benchmarks buyers care about, the 40% to 60% price gap Sakana claims comes straight out of somebody’s gross margin.
The winners are less obvious. Open-weight model providers gain demand as workers inside the pool without needing a sales team. NVIDIA gains too — its Nemotron family entered the pool through an August 2026 partnership.
This is the same pressure that produced DeepSeek V4.1 Flash and Cognition’s SWE-2 on a Chinese base. The difference is that Sakana is not competing on training. It is competing on allocation.
Sakana raised roughly $200 million at about a $2.7 billion post-money valuation in its Series B announced November 17, 2025, backed by MUFG, Khosla Ventures, In-Q-Tel and Macquarie Capital. That is a fraction of what the labs it routes around have raised.
Which is the point of the business model. Orchestration does not need a $45 billion data center contract behind it.
What are the catches?
The $2 headline is not the bill. Orchestration tokens are added on top of the input and output you can see, so a multi-step task costs more than a naive token count suggests. Several other limits land on real workloads.
- Hidden orchestration billing. Coordination tokens add to the total beyond visible input and output, per AiCybr’s teardown.
- The 272K cliff. Ultra v2 input doubles to $10 per million above 272,000 tokens. Long-context users should model this before migrating.
- No weights, no self-hosting. Hosted API only. Air-gapped and regulated deployments are out.
- No EU or EEA availability. European buyers cannot use it at launch.
- Silent parameter quirks. Temperature is accepted but ignored, and reasoning.effort values “xhigh” and “max” currently map to the same behavior.
- Self-reported benchmarks. No independent evaluation has been published yet.
- Tool-call costs compound. At $0.007 per web call, an agent loop can outspend its own token bill.
The one that should bother finance teams most is the first. A router whose cost depends on how many hops it chooses is a variable-cost product sold with a fixed-price headline.
Frequently asked questions
When did Sakana Fugu Max launch?
September 11, 2026, alongside Fugu Ultra v2.0. Both went live immediately through an OpenAI-compatible API and on OpenRouter.
Is Fugu Max a foundation model?
No. It is a learned orchestrator that routes subtasks across a pool of other models. The answers you get are produced by workers Sakana selects at runtime.
What is the context window?
1 million tokens on both tiers, with a 128,000-token maximum output. Ultra v2 charges a long-context surcharge above 272,000 tokens.
Can I download the weights?
No. Fugu is hosted only. If self-hosting matters, look at open-weight releases instead.
How does the price compare to frontier APIs?
Sakana claims Fugu Max costs 40% to 60% less than Sonnet 5, GPT-5.6 Terra and Kimi K3. At $2 per million input, it undercuts the $10 tier that frontier models charge.
Which models are in the pool?
Open-weight and specialized models, including NVIDIA’s Nemotron family. Fable 5, Fable 5.1 and GPT-6-Astra were excluded from Ultra v2’s pool.
Is it available in Europe?
Not at launch. Sakana lists Fugu as unavailable in the EU and EEA.
The bottom line
Sakana Fugu Max is the first credible attempt to make model selection a product rather than a decision. At $2 per million input tokens against a claimed 40% to 60% discount, the pricing is aggressive enough to force a response from anyone selling mid-tier inference.
The verdict: worth piloting, not worth standardizing on yet. Run it against a workload you already price, measure the real bill including orchestration tokens, and see whether the 40% to 60% claim survives contact with your traffic.
The strategic read matters more than the benchmark table. If routing works, the value in this market shifts from who trains the best model to who allocates across them — and that is a far cheaper business to run than a $2.7 billion valuation implies it should be. Compare it against the field in our roundup of the cheapest 1M context models before you commit.
Leave a Reply