Sakana Fugu isn't a new foundation model. It's a multi-agent orchestration layer sitting behind one OpenAI-compatible endpoint (model ID fugu-ultra-20260615), routing your request to whichever underlying frontier model handles it best. That distinction matters for pricing: the sticker price per token isn't your real cost, because Fugu also bills for the coordination work happening behind the scenes. If you're evaluating it for production, budget for orchestration overhead separately from the quoted rate.
What Fugu actually is: an orchestration layer, not a model
Sakana AI was founded by Llion Jones, a co-author of "Attention Is All You Need," and David Ha, former Google Brain research director. They named the product after the idea that a school of fish (*sakana*) outperforms any single fish, and a fugu (pufferfish) is the one species that has to get every move right or it's fatal. The metaphor is the pitch: instead of betting everything on one model's weaknesses, Fugu coordinates several.
There are two tiers. Fugu is the everyday version for coding and chat-style tasks. Fugu Ultra is built for harder, multi-step work, the kind of task where a single model tends to drift or miss context across a long session. Under the hood, the aggregation layer does contradiction detection, confidence weighting, chain-of-thought retention, and domain-aware routing, then hands you one answer instead of several to reconcile yourself.
Pricing: the number on the page isn't the number you'll pay
Here's Fugu Ultra's published token pricing, verified against Sakana's official page:
| Tier | Input | Output | Cached input |
|---|---|---|---|
| Standard context (≤272K tokens) | $5 / M tokens | $30 / M tokens | $0.50 / M tokens |
| Extended context (>272K tokens) | $10 / M tokens | $45 / M tokens | $1.00 / M tokens |
Subscription plans run $20/month (Standard), $100/month (Pro, 10x the usage), and $200/month (Max, 20x the usage). Sign up before the end of July 2026 and Sakana is currently offering a free second month on any tier, worth locking in now if you're planning to test it this quarter.
Base Fugu (non-Ultra) prices differently depending on how many agents get involved: when a single agent handles the task, you pay that underlying model's standard rate; when multiple agents coordinate, Sakana says it never stacks per-model fees on top of each other — you're billed at the rate of the highest-tier model used, not the sum of all of them.
The orchestration cost nobody prices out for you
The pricing pages don't spell this out clearly, so here's the math. Say you send Fugu Ultra a task that would cost $0.03 in a single GPT-5.5 call (10K input tokens, 2K output tokens, at GPT-5.5's own rates). Fugu Ultra doesn't run that once; it may dispatch the same prompt-equivalent work across 2-3 candidate models internally, run a contradiction check, and only then return the reconciled answer. If that adds even one extra full round of input+output tokens at Fugu Ultra's rate, you're looking at roughly $0.08-$0.12 for what looked like a $0.03 job on the label. On a single request that's noise. Run it across a few thousand production calls a month and the gap between the quoted rate and the actual bill becomes the line item that blows your budget forecast.
The practical takeaway: don't estimate Fugu Ultra costs from the per-token rate alone. Track completed-task cost during a pilot — actual dollars per resolved ticket or merged PR — not tokens per call, and compare that against what your current single-model setup costs for the same volume of finished work.
Benchmarks: Fugu Ultra vs Claude Fable 5 vs GPT-5.5
| Benchmark | Fugu Ultra | Fugu | Claude Fable 5 | GPT-5.5 |
|---|---|---|---|---|
| SWE-Bench Pro | 73.7 | 59.0 | — | — |
| LiveCodeBench | 93.2 | 92.9 | 89.8 | 85.3 |
| TerminalBench 2.1 | 82.1 | 80.2 | — | — |
| GPQA-Diamond | 95.5 | 95.5 | — | — |
| Humanity's Last Exam | 53.3 | — | in the 44.3-53.3 range | in the 44.3-53.3 range |
Sakana's own framing — and early third-party reporting backs this up — is that the gap widens in long, messy, real-world workflows rather than isolated benchmark runs. On code review specifically, Fugu Ultra has been reported to catch bugs GPT-5.5 misses and give more complete answers, which lines up with what you'd expect from a system that cross-checks multiple models' outputs before answering instead of committing to one pass. If you want the fuller picture on where GPT-5.5 stands against the newer GPT-5.6 tier, or how Claude Fable 5 compares to Anthropic's own Mythos line, those breakdowns go deeper on each model individually.
How to access it: direct API, or through an aggregator
Fugu ships as an OpenAI-compatible API, and Sakana's own page lists official support for OpenRouter, Vercel, OpenCode, Creao, and Merge — so if you're already routing traffic through one of those, adding Fugu is a config change, not a new integration.
Two things to know before you connect directly to Sakana:
- EU/EEA access isn't available yet. Sakana states this directly on its own pricing page, framed as ongoing compliance work with no stated timeline. There's no confirmed workaround; whether a given aggregator can route around the restriction depends on where that aggregator's infrastructure sits, and Sakana doesn't publish anything on that either, so don't assume it's solved just because you're going through a third party.
- Billing runs through Sakana's own dashboard, separate from whatever you're already using for Claude, GPT, or Gemini calls — one more invoice to reconcile if you're already juggling several model providers.
Because Fugu is OpenAI-compatible, adding it to an existing setup is usually a matter of swapping the base URL and model string in whatever client you're already using; nothing else in your request/response handling needs to change. That's also why aggregators like AIReiter that already sit in front of Claude, GPT, and Gemini can add a new model without asking you to touch your integration code.
Should you use it?
Fugu Ultra makes sense when accuracy matters more than latency or cost predictability — long-running coding tasks, research workflows where you'd otherwise manually cross-check two or three models, or anything close to the edge of what a single model can reliably do alone.
Skip it when you need real-time responses, the task is simple enough that one good model already nails it, or your budget is fixed and can't absorb the orchestration variance described above. For high-volume, low-complexity calls, a single well-chosen model will almost always be cheaper and more predictable.
FAQ
Is Sakana Fugu free?
No. Fugu Ultra is priced at $5/M input and $30/M output tokens (standard context), or via subscription starting at $20/month. Sakana is currently running a free-second-month promotion for signups before the end of July 2026.
Is Sakana Fugu open source?
Sakana hasn't published Fugu's weights or orchestration code as open source. It's accessed exclusively through its hosted OpenAI-compatible API.
Is Fugu available on OpenRouter?
Yes, it's one of five platforms Sakana officially lists (alongside Vercel, OpenCode, Creao, and Merge), so you don't need a direct account with Sakana to try it if you're already set up on any of those.
What does "Fugu" mean in this context?
Fugu is pufferfish, a dish that's harmless when prepared correctly and dangerous when it isn't. Sakana picked the name specifically for that all-or-nothing property: the whole design goal of the orchestration layer is to catch the one mistake a single model would have let through.
Bottom line
The benchmark gains are real, and they show up more in messy multi-step work than in isolated test runs. What most coverage of Fugu misses is that the number on the pricing page isn't what you'll actually be billed once orchestration overhead is added in, so treat the quoted rate as a floor, not an estimate, and pilot on completed-task cost before you commit budget to it.