What is Sakana Fugu (Fugu, Max and Ultra v2)?
Sakana Fugu is a language model from Tokyo-based Sakana AI, generally available since 22 June 2026, that works as a whole multi-agent system behind one API. Rather than answering every request itself, it decides whether to delegate parts of a task to other models in a pool, including copies of itself, then checks and combines their work. It builds on Sakana's TRINITY and Conductor research papers.
It comes in versions. Base Fugu aims at low latency for everyday work. Fugu Ultra, updated to v2.0 on 11 September with a 1-million-token context and 128K output, aims at the hardest problems. Fugu Max, added the same day, uses a pool of open-weight and specialist models to balance cost and quality. Fugu Cyber is sold separately through sales. Sakana does not reveal which models handle a request.
Prices per million tokens are $2 input and $6 output for Fugu Max and $5 and $30 for Ultra v2, rising to $10 and $45 above 272K tokens of context. Base Fugu charges the rate of the most expensive model it uses. Subscriptions cost $20, $100 or $200 a month. It is not available in the EU or EEA.
How much does Sakana Fugu (Fugu, Max and Ultra v2) cost?
| Plan | Price | What you get |
|---|---|---|
| Fugu Max | $2 in / $6 out per 1M tokens | Cached input $0.25; web_search and web_fetch $0.007 per call |
| Fugu Ultra v2 | $5 in / $30 out per 1M tokens | $10 / $45 above 272K context; cached $0.50 |
| Fugu | Rate of the top model used | Fees for several agents are not stacked |
| Subscriptions | $20, $100 or $200 a month | Standard, Pro (10x) and Max (20x); all include Fugu, Ultra and Max |
Fugu Max $2 input / $6 output per 1M tokens ($0.25 cached); Fugu Ultra v2 $5 / $30 ($0.50 cached), $10 / $45 above 272K context; base Fugu at the rate of the top model used. Subscriptions $20, $100 or $200 a month. Web tools $0.007 per call.
Is Sakana Fugu (Fugu, Max and Ultra v2) worth it?
Fugu is a different kind of product: a model that manages other models. On Sakana's launch charts, Fugu Ultra scored 73.7 on SWE-Bench Pro against 69.2 for Claude Opus 4.8 and 95.5 on GPQA-Diamond, while trailing Claude Fable 5 on SWE-Bench Pro (80.0) and Humanity's Last Exam. The September Ultra v2 reports 74.3 on DeepSWE. None of this has been run independently, and Sakana's internal SWEFish test features in several wins.
Its real argument is resilience. Sakana pitches it as a hedge against export controls and lost access: if a lab restricts a model, Fugu routes around it. For a Japanese or other non-US company worried about access to US frontier models, that is a genuine selling point.
The costs are opacity and unpredictability. Sakana says routing information is "not exposed by design", so you cannot tell which providers saw a prompt, and only base Fugu lets you exclude models; Ultra and Max use fixed pools. Orchestration tokens are billed on top of the visible prompt and answer. In one independent test reported by VentureBeat, Fugu Ultra built a game in 22 minutes for about $7, against 79 minutes and $38 for Opus 4.8, but Opus's result was better.
Training on your data is on by default, with an opt-out in the console, and the service is closed to the EU and EEA. Fugu Max at $2/$6 is the version most teams should try first.
Where it earns its place
A Japanese company worried about losing access to US models
Keep frontier-level AI working if a provider restricts access
Fugu routes around restricted or unavailable models automatically. Start with a $20 Standard subscription to test, then move to pay-as-you-go for production priority.
A research team running long, multi-step analyses
Literature reviews, paper reproduction and patent searches
Fugu Ultra v2 is tuned for these, with a 1M-token context. A 200K-token task with a 20K-token report costs about $1.60 before orchestration overhead.
A developer team doing code review
Find more issues per pull request
Base Fugu targets code review and Codex-style coding, and lets you exclude providers your policy bars. One early user said it surfaced more than twenty issues where other tools flagged about three.
Before you commit
- Setup
- Sign up at console.sakana.ai, choose a subscription or pay-as-you-go, and call fugu-max-v1.0 or fugu-ultra-v2.0 through the OpenAI-compatible API; also on OpenRouter. Turn off training in the console first.
- Integrations
- OpenAI-compatible API; OpenRouter and Vercel AI Gateway listings; guides exist for using it with Claude Code. Built-in web_search and web_fetch tools.
- Data and privacy
- Sakana's FAQ says customer data may be used to improve Fugu, with an opt-out in the console at any time. Retention is not stated. Routing is hidden, so which outside providers saw a prompt is unknown. Not offered in the EU or EEA pending GDPR compliance.
- Cost at scale
- Fugu Max: 1,000 agent tasks of 500K input and 30K output tokens cost about $1,180 before orchestration tokens; Ultra v2 at over 272K context about $6,350. Base Fugu bills at the top model's rate.
- Limits
- Ultra v2: 1M context, 128K output. Routing and pool contents are hidden; only base Fugu allows excluding models. Not available in the EU or EEA.
What's good
- Routes around providers that restrict access, by design
- Fugu Max at $2/$6 per million tokens tops several of Sakana's tests
- One OpenAI-compatible API for a pool of frontier models
- Subscriptions from $20 a month include all three versions
What's not
- Which models handled a request is never shown
- Ultra and Max pools cannot be changed
- Trains on customer data unless you opt out; not available in the EU or EEA
- Orchestration tokens make bills hard to predict
Best for
Teams outside the EU that want near-frontier results for multi-step research and coding without depending on a single model provider.
Sakana Fugu (Fugu, Max and Ultra v2) alternatives
Sakana Fugu (Fugu, Max and Ultra v2): common questions
- Is Sakana Fugu (Fugu, Max and Ultra v2) worth it?
- A clever way to get near-frontier answers without depending on one lab, but you cannot see which models handled your data, and the best scores are Sakana's own. We rate it 3.9 out of 5.
- How much does Sakana Fugu (Fugu, Max and Ultra v2) cost?
- Fugu Max $2 input / $6 output per 1M tokens ($0.25 cached); Fugu Ultra v2 $5 / $30 ($0.50 cached), $10 / $45 above 272K context; base Fugu at the rate of the top model used. Subscriptions $20, $100 or $200 a month. Web tools $0.007 per call.
- What are the downsides of Sakana Fugu (Fugu, Max and Ultra v2)?
- Which models handled a request is never shown; ultra and Max pools cannot be changed; trains on customer data unless you opt out; not available in the EU or EEA; orchestration tokens make bills hard to predict.
- Who is Sakana Fugu (Fugu, Max and Ultra v2) best for?
- Teams outside the EU that want near-frontier results for multi-step research and coding without depending on a single model provider.
- How much does Sakana Fugu (Fugu, Max and Ultra v2) cost for a team or in production?
- Fugu Max: 1,000 agent tasks of 500K input and 30K output tokens cost about $1,180 before orchestration tokens; Ultra v2 at over 272K context about $6,350. Base Fugu bills at the top model's rate.
- Is Sakana Fugu (Fugu, Max and Ultra v2) safe for company data?
- Sakana's FAQ says customer data may be used to improve Fugu, with an opt-out in the console at any time. Retention is not stated. Routing is hidden, so which outside providers saw a prompt is unknown. Not offered in the EU or EEA pending GDPR compliance.
- What are the best Sakana Fugu (Fugu, Max and Ultra v2) alternatives?
- The closest alternatives we have reviewed are Claude Opus 5.5, GPT-6.1 Sol, MiMo-V2.6-Pro and Kimi K3.1 Pro.
Answers checked Oct 25, 2026.