Sakana priced Fugu Max 60% under the frontier — but the meter counts the agents talking to each other, and that is 5-12x the tokens you see
TL;DR: On 11 September 2026 Sakana AI split its Fugu orchestrator into two products. Fugu Max: $2 input / $6 output / $0.25 cached per million tokens, flat. Fugu Ultra v2: $5 / $30 / $0.50, rising to $10 / $45 / $1.00 above 272K context. Tool calls $0.007 each. The pitch is a 40-60% undercut on Sonnet 5, GPT-5.6 Terra and Kimi K3 output, plus best-or-joint-best scores on eleven benchmarks across the two models. The number that decides your bill is not on the rate card: Fugu is a router that bills the agents’ conversation with each other, at standard rates, and independent measurement of Fugu Ultra v1 put that overhead at 5x to 12x the visible output — a fixed ~1,260-token floor per request, then 20,487 orchestration tokens for a 2,223-token answer. Run that through Max’s rates and the effective cost lands near $24-$61 per million useful output tokens against a $6 sticker; Opus 5 is $25 with no multiplier. Sakana reports the overhead honestly in the usage object. It has not published what to expect. For buyers: measure your own ratio before you migrate — and note Fugu is unavailable in the EU/EEA and routing is undisclosed by design.
The product is the routing, and the routing is the invoice
Most model launches ask you to compare two rate cards. This one does not work that way, and the difference is the whole story.
Fugu is not a model. It is a learned orchestrator that takes your request, assembles a task-specific scaffold of other models — open-weights systems, specialist models, and NVIDIA’s Nemotron family through a collaboration announced with the release — assigns them roles, has them delegate and critique and verify, then synthesises an answer. Sakana’s published research describes the mechanics: TRINITY assigns Thinker, Worker and Verifier roles; the Conductor uses reinforcement learning to discover coordination strategies in natural language.
That is a genuinely interesting piece of engineering, and the benchmark sheet is not trivial. Fugu Max claims best overall on six benchmarks including Terminal Bench 2.1, GPQA Diamond, AA-LCR, GDP.pdf, AutomationBench and SWEFish. Fugu Ultra v2 claims best or joint-best on five, including 48.3 on Chartography against Opus 5’s 27.3 and Fable 5’s 29.5, and 74.3 on DeepSWE.
But an orchestrator has a property a model does not. When you buy a model, the tokens you are billed for are approximately the tokens you can read. When you buy an orchestrator that charges per token, you are billed for a conversation you never see.
What the rate card says
| Input | Output | Cached input | |
|---|---|---|---|
| Fugu Max | $2.00 | $6.00 | $0.25 |
| Fugu Ultra v2 (≤272K) | $5.00 | $30.00 | $0.50 |
| Fugu Ultra v2 (>272K) | $10.00 | $45.00 | $1.00 |
| Claude Opus 5 | $5.00 | $25.00 | — |
| GPT-6 Astra (≤272K) | $10.00 | $50.00 | — |
| Claude Fable 5.1 | $10.00 | $50.00 | $0.25 |
Per million tokens. Tool calls — web_search, web_fetch — bill at $0.007 per call on top.
Two things worth pausing on before the arithmetic. First, Fugu Ultra v2’s output rate of $30 is above Opus 5’s $25, so the “undercut” claim belongs to Max, not to Ultra. Second, Sakana has adopted the same 272K context threshold that GPT-6 Astra uses to dual-price a request. That boundary is quietly becoming an industry convention, and it has the same trap: crossing it reprices the request, so a long agent session that drifts over the line gets more expensive in a step rather than a slope.
The meter nobody quotes
Here is the disclosure that should reframe the launch. Fugu’s billing does not hide orchestration work — it separates it. The API returns orchestration_input_tokens and orchestration_output_tokens alongside your visible usage, and those tokens are billed at the same standard rates, with no additional per-agent fee.
That is more honest than the alternative. It is also a meter with no published expectation attached to it.
Requesty’s reverse-engineering of Fugu Ultra v1 is the only public measurement of the size of that overhead:
| Task | Visible tokens | Orchestration | Total | Multiple | Latency |
|---|---|---|---|---|---|
| ”What is 2+2?“ | 262 | 1,260 | 1,522 | 5.8x | 8s |
| ”Explain recursion” | 622 | 1,260 | 1,882 | 3.0x | 13s |
| Math proof | 1,713 | 18,794 | 20,507 | 12.0x | 101s |
| ”Compare Python vs Rust” | 2,223 | 20,487 | 22,710 | 10.2x | 156s |
Note the shape. There is a fixed floor of roughly 1,260 orchestration input tokens on every request — the system prompt injected into worker models — which dominates trivial queries. Then, on anything requiring real coordination, overhead jumps by an order of magnitude. The multiplier is not a constant you can budget with. It is a function of how hard your question is, decided by a policy you cannot inspect.
Running it through Max’s rates
Sakana does not publish how orchestration overhead splits between input and output, so the honest result is a range rather than a figure. Take the Python-versus-Rust request — 2,223 visible output tokens, 20,487 orchestration tokens — and price it on Fugu Max:
- If every orchestration token billed as input at $2/M: ≈$0.054 per request
- If every orchestration token billed as output at $6/M: ≈$0.136 per request
Divide by the output you can actually use:
| Sticker output rate | Effective, per M visible output | |
|---|---|---|
| Fugu Max — complex query | $6.00 | ~$24 to $61 |
| Fugu Max — trivial query | $6.00 | ~$16 to $35 |
| Fugu Ultra v2 — complex query | $30.00 | ~$76 to $306 |
| Claude Opus 5 | $25.00 | $25 |
The 60% undercut survives contact with short answers. On the reasoning work Fugu is actually sold for, it lands somewhere between parity with Opus 5 and several times worse — and the range is wide because the split is unpublished, not because the estimate is sloppy.
Two caveats that genuinely favour Sakana. These measurements are of Fugu Ultra v1, not of Max or of v2. And v2’s most interesting design choice is that it removed Claude Fable 5, Fable 5.1 and GPT-6 Astra from its pool and scored higher anyway — a smaller pool coordinating better could plausibly coordinate more cheaply too.
Neither caveat rescues the rate card. They just relocate the problem: the overhead figure is the single most important number for pricing this product, and it is the one number Sakana has not published for the products it launched today.
Why the opacity is the risk, not the price
Routing information is “not exposed by design” — Sakana’s own words. You get token totals without attribution.
That is defensible as trade secret and uncomfortable as procurement. This desk has now covered the same structural hazard three times this month from different directions. DeepSeek made capability downgrade explicit by routing reasoning away at peak hours. CISA’s distillation advisory described degraded models served with enough variation that evaluation cannot detect the substitution. And when OpenAI shipped a free harness that decides how many tokens you spend, the problem was that the spend controller belonged to the counterparty.
Fugu is not accused of any of that. It has simply built a product whose entire value is a substitution decision, and made that decision unobservable. A routing policy that quietly prefers a cheaper worker is indistinguishable from a good day, and your evaluation suite will not flag it.
There are real mitigations, and they are better than nothing: you can opt out of specific providers or models in console settings, and opt out of training-data use at any time. A veto is not visibility, but it is a lever.
One more reason to sample rather than trust the launch sheet: the benchmark claims are unreproduced. A 21-point margin over Opus 5 on Chartography is extraordinary, and extraordinary margins are exactly what private test sets and repriced index runs exist to check.
The EU line, again
Sakana’s product page states Fugu is not yet available in the EU or EEA pending GDPR compliance. No timeline.
That completes an unwelcome pattern for European buyers in a single quarter: GPT-6 Astra without EU data residency, the Agents API US-only with no zero-data-retention, and now the quarter’s most novel orchestration product unavailable in the bloc outright. If you are procuring in Europe, the multi-agent shelf this quarter holds self-hosted harnesses and very little else.
What to do with this
- Measure your own overhead ratio before migrating anything. Sakana gives you the instrument: sum
orchestration_input_tokensandorchestration_output_tokensacross fifty real requests, divide by visible output, and compute effective cost per useful token. Compare that to your incumbent — never the rate card. - Split your evaluation by task length. Short, high-volume, latency-tolerant calls are where Fugu Max’s pricing most plausibly wins. Long reasoning chains are where the multiplier eats the discount.
- Rule it out of interactive paths on latency alone. 101 to 156 seconds on reasoning tasks is a batch product. Price it against agent-hour economics, not against chat.
- Watch the 272K line on Ultra v2, for the same reason it matters on Astra: crossing it reprices the whole request in a step.
- Sample outputs over time, not just at pilot. Undisclosed routing means a quality change arrives without a changelog. Keep a fixed prompt set and re-run it monthly.
- Pilot cheaply, commit slowly. The OpenAI-compatible API and one-line parameter change make trying Fugu nearly free — which is a reason to try it and a poor reason to standardise on it while the overhead figures are unpublished and the benchmarks unreproduced.
- If you are in the EU, price the alternative. A frontier model with aggressive cache discounts — Fable 5.1’s cache-read cut to $0.25 is the relevant one — plus your own harness is the available option, not a worse version of an unavailable one.
For the broader field, the best AI agent tools rundown covers orchestration platforms and best AI harnesses covers the loop layer Fugu competes with. The Gemini review is relevant here in an unusual way — Requesty’s probing identified Fugu Ultra v1’s workers as Gemini-backed — while the Claude review tracks the Opus and Fable tiers this pricing is measured against.
Frequently asked questions
What do Fugu Max and Fugu Ultra v2 actually cost?
Fugu Max is $2 per million input tokens, $6 per million output, and $0.25 per million cached input — flat, with no context-length tier. Fugu Ultra v2 is $5 input, $30 output and $0.50 cached below 272K tokens of context, rising to $10 input, $45 output and $1.00 cached above it. Tool calls are billed separately at $0.007 per call for web_search and web_fetch. Those are the sticker rates, and for a normal single model they would be the whole story. Fugu is not a single model — it is an orchestrator that delegates to a pool of other models — and its billing documentation is explicit that the tokens consumed by that internal coordination are real usage, counted in the final price at the same standard input and output rates. So the rate card tells you the price per token and says nothing about how many tokens a request will take. For Fugu that second number is the one with the variance in it.
What are orchestration tokens and why do they change the arithmetic?
They are the tokens Fugu spends talking to itself. When a request comes in, Fugu assigns roles across a pool of models — the published TRINITY research describes Thinker, Worker and Verifier assignments — delegates sub-tasks, has agents check each other's work, and synthesises a final answer. Every delegation is an input payload to another model and every critique is an output from one. Sakana reports these separately in the usage object as orchestration_input_tokens and orchestration_output_tokens, which is more disclosure than most routers offer, and bills them at the same rates as your visible tokens with no stacking fee per agent. The reason this matters is scale. Independent reverse-engineering of Fugu Ultra v1 by Requesty measured a fixed floor of roughly 1,260 orchestration input tokens injected into worker models on every request regardless of complexity, then found total consumption running 5x to 12x the visible output: 1,522 total tokens for a 262-token answer to 'What is 2+2?', and 22,710 total for a 2,223-token comparison of Python and Rust. A 10x token multiplier applied to a 60% price cut is not a 60% price cut.
So is Fugu Max actually cheaper than Opus 5 or Sonnet 5?
For short answers, plausibly yes. For long reasoning work, the undercut can disappear entirely — and Sakana has not published the figures that would settle it. Work the published v1 measurement through Fugu Max's rate card. The Python-versus-Rust request produced 2,223 visible output tokens against 20,487 orchestration tokens. Sakana does not publish how that overhead splits between input and output, so the honest answer is a range: if all of it billed as input at $2 per million the request costs about $0.054, and if all of it billed as output at $6 per million it costs about $0.136. Divide by the useful output and the effective rate is roughly $24 to $61 per million visible output tokens, against a $6 sticker. Claude Opus 5 charges $25 per million output tokens with no multiplier. On the trivial query the same arithmetic gives $16 to $35 effective, still several times sticker. Two caveats cut in Sakana's favour: those measurements are of Fugu Ultra v1, not of Max or v2, and v2's deliberately smaller pool may well coordinate more cheaply. The point is not that Fugu is expensive. It is that nobody outside Sakana currently knows, and the rate card is not evidence.
Can I see which model answered my request?
No, and Sakana says so plainly rather than burying it — the product documentation states that the models Fugu selects and how it coordinates them are proprietary, so routing information is not exposed by design. You get aggregate token accounting without attribution. There are two partial mitigations. You can opt out of specific providers or models in console settings, which gives you a veto without giving you visibility, and you can opt out of training-data use at any time. The residual risk is one this desk has flagged repeatedly in a different guise: when a vendor controls substitution and you cannot observe it, your evaluation harness cannot detect a quality change, and a routing policy that silently reaches for a cheaper worker looks identical to a good day. DeepSeek made that trade explicit by downgrading reasoning at peak hours, and CISA's distillation advisory described degraded models served in a way evaluation is not meant to catch. Fugu is not accused of either. It simply has the same unobservable surface, so the burden falls on you to sample outputs over time rather than trust a launch benchmark.
Can EU buyers use this at all?
Not today. Sakana's own product page states that Fugu is not yet available in the EU or EEA while the company works toward GDPR compliance. There is no published timeline. For a European buyer that is a hard blocker rather than a procurement footnote, and it is worth noticing how ordinary this has become: GPT-6 Astra launched without EU data residency, the Agents API went to public beta US-only with zero data retention unsupported, and now the most interesting orchestration product of the quarter is unavailable in the bloc entirely. The practical consequence is that EU teams evaluating multi-agent orchestration this quarter are choosing between self-hosted harnesses and nothing. If you are in that position, price your own orchestration on a frontier model with cache discounts and treat the Fugu comparison as a benchmark you cannot yet buy.
What should I actually do before committing spend to Fugu?
Measure your own overhead ratio before you migrate anything, because Sakana has handed you the instrument to do it. The usage object breaks out orchestration_input_tokens and orchestration_output_tokens on every response. Run fifty representative requests from your real workload, sum those fields against your visible output, and compute the multiplier and the resulting effective cost per useful token. That number, not the rate card, is what you compare to your incumbent. Two things to watch while you do it. Latency is a product constraint here, not a footnote — the v1 measurements recorded 8 seconds for a trivial query, 101 seconds for a math proof and 156 seconds for a comparison task, which rules Fugu out of interactive paths regardless of price. And the benchmark claims remain unreproduced: the 48.3 on Chartography against Opus 5's 27.3 is a 21-point margin with no third-party replication, and the site's coverage of private test sets is a reminder that vendor-run evaluations and buyer-run ones diverge. Migration cost is genuinely low — the API is OpenAI-compatible and Sakana documents a single-line parameter change — which is an argument for piloting it and a bad argument for committing to it.
Sources
- Sakana AI — Introducing Fugu Max and Fugu Ultra v2: Orchestrating the Pareto Frontier (11 September 2026)
- Sakana AI — Fugu: Multi-Agent System as A Model (product page, routing disclosure and EU/EEA availability)
- Sakana AI Console — Pricing (Fugu Max and Fugu Ultra v2 token rates, 272K tier, tool call rates)
- Requesty — Inside Sakana Fugu Ultra: We Reverse Engineered Its Multi-Agent Architecture (orchestration token measurements)
- MarkTechPost — Sakana AI launches Fugu Max and Fugu Ultra v2 for cheaper, stronger multi-agent orchestration (10 September 2026)
- OrcaRouter — Sakana Fugu Ultra v2: a smaller pool, a higher score (pool exclusions and benchmark caveats)
Related tool reviews
Questions or corrections? Email Pick Right. Want the full list? See all news.