Published 11 September 2026
Sakana AI released Fugu Max and Fugu Ultra v2 on 11 September 2026. Fugu Max costs $2 per million input tokens and $6 per million output tokens, which Sakana says is 40 to 60 percent cheaper on output than Sonnet 5, GPT 5.6 Terra and Kimi K3. Both models run as a hosted API only, and neither is available in the EU or the EEA.
What are Fugu Max and Fugu Ultra v2?
Fugu is not a single model. It is an orchestrator, a model trained to hand each task to other models and stitch the answers back together. You send one request to one API, and Fugu decides behind the scenes which models in its pool should do the work.
The two new versions share that architecture and differ only in what they aim at. Fugu Max chases the best result per dollar. It widens the pool with a large set of open-weight and specialist models, including NVIDIA’s Nemotron family, and routes each task to the cheapest model that can still solve it. Fugu Ultra v2 goes the other way and chases the best possible answer on hard, multi-step work.
Sakana makes a second argument that has nothing to do with quality. Because the pool is swappable, you are not tied to one supplier. Sakana names vendor lock-in, revoked API access and sudden service cutoffs as the risks this design is meant to remove.
Fugu Max and Fugu Ultra v2 pricing and specifications
The table below shows where the two models separate. The architecture is the same, so price is the real difference.
| What is measured | Fugu Max | Fugu Ultra v2 |
|---|---|---|
| Input price per 1M tokens | $2 | $5 |
| Output price per 1M tokens | $6 | $30 |
| Cached input per 1M tokens | $0.25 | not published for v2 |
| Context window | 1 million tokens | 1 million tokens |
| Maximum output | 128,000 tokens | 128,000 tokens |
| Long-context surcharge | none, flat rate | yes, above 272,000 tokens |
These are list prices from Sakana’s release post and its OpenRouter model pages. The $0.25 cached-input rate for Fugu Max comes from OpenRouter’s provider table rather than from Sakana. Sakana’s own pricing page still shows the older Fugu Ultra v1.1 rates, so the size of the long-context surcharge on v2 is not yet public.
On benchmarks, Sakana reports the following:
- Chartography, which tests reading charts and structured data: Fugu Ultra v2 scores 48.3, against 27.3 for Opus 5 and 29.5 for Fable 5
- DeepSWE, a real-world software engineering test: Fugu Ultra v2 scores 74.3, beating models that cost three to five times more per token
- Fugu Ultra v2 takes the best or joint-best score on five of eight benchmarks, and a top-two place on seven of eight
- Fugu Max takes the best overall score on six benchmarks, including Terminal Bench 2.1, GPQA Diamond and AutomationBench
- Fugu Max improves on the cost-performance frontier in seven of ten benchmarks
These are Sakana’s own evaluations. Its June release note said the rival scores came from the model providers themselves; the new post does not repeat that statement, so it is unclear where the Opus 5 and Fable 5 figures were taken from.
What Sakana AI is not saying about Fugu Max and Fugu Ultra v2
Only two benchmark numbers appear as text in the announcement. Every other claim is phrased as a count of wins, such as best on six benchmarks, with the actual scores sitting inside chart images. You cannot check the margin on most of them.
One of the benchmarks Sakana counts is its own. SWEFish is built from Sakana’s internal coding challenges, has not been released, and contributes a win to both models’ totals.
Sakana also does not say which models sit in the pool, and its own FAQ confirms that the routing is hidden by design. You cannot see which model answered your prompt, and the pool can change without you knowing. For a buyer who has to document where data went, that is the awkward part of the vendor-independence pitch.
There is no speed figure in the announcement either. OpenRouter measures Fugu Max at 13 tokens per second, with a first response after 5.37 seconds. That is an outside measurement on a model hours old, but it is the only timing number available, and it is slow for anything interactive.
What you get on day one
- Access: hosted API only, OpenAI-compatible, with no weights to download and run yourself
- Availability: not offered in the EU or the EEA while Sakana works on GDPR compliance
- Inputs: text, images and PDFs, with web search built in
- Controls: reasoning effort set to high, xhigh or max, plus function calling and structured outputs
- Switching: existing Fugu users change one parameter, with no migration
- Training cutoff: 28 August 2026 for Fugu Ultra v2
What this means
For a Dutch or European team, this release is safe to ignore today, whatever the benchmarks say. Sakana does not serve the EU or the EEA, so you cannot put either model into production here. Outside Europe it is a different story. Picture a ten-person platform team pushing code review or document extraction through a model all day, where the monthly bill is the real constraint. For that team, Fugu Max at $6 per million output tokens deserves a proper test against whatever they run now.
The claim worth arguing with is the supply-chain one. Sakana says a swappable pool protects you from any single vendor. That holds right up to the point where you notice you have swapped a known dependency for an unknown one: you cannot see the pool, you cannot audit which model read your data, and Sakana can change both without telling you. The first thing we would measure is not a benchmark. Run the same workload through Fugu Max and through the single cheapest model you would otherwise pick, and compare cost per finished task rather than cost per token. Orchestration tokens are billed as ordinary tokens, and a task that quietly consults four models is where the price advantage can disappear.
For more information, visit the official announcement of Fugu Max and Fugu Ultra v2 on the Sakana AI blog.
