Fugu suits engineering teams that want automatic failover across multiple frontier AI vendors instead of committing to one, especially anyone caught by the June 2026 export-control gap that blocked two of Anthropic's own models. Hacker News testers called Fugu's own output quality well short of those same blocked models, so pilot it before replacing a working single-vendor setup.
Fugu is a multi-agent orchestration model from Sakana AI that scores 95.5% on GPQA Diamond by routing each request through a learned coordinator instead of one trained network. Released in June 2026, it exposes three pooled frontier LLMs as a single OpenAI-compatible API endpoint, so callers never choose the underlying model themselves.
Where it sits
- $11.25/M$ per 1M tokensBlended price (3:1)Lower is better#55 / 64peer median $1.70/Mvendor price, checked by HokAI
- --tokens/sOutput speedHigher is better-- / 39peer median 90 tok/scited: Artificial Analysis
- --% solvedSWE-bench VerifiedHigher is better-- / 28peer median 78.3%per source, see benchmark scores
- 95.5%% correctGPQA DiamondHigher is better#2 / 44peer median 88.3%per source, see benchmark scores
Pricier than 90% of the 64 GA models with a published price, and in the top third on GPQA Diamond (rank 2 of 44).
Ranks are against GA models on HokAI that publish the same figure; ties share a rank.
Provider: Sakana AI · Family: Fugu
Input modalities: text, tool-calls · Output: text, tool-calls
About Fugu
Fugu is a multi-agent orchestration system built by Sakana AI, the Tokyo research lab founded in July 2023 by David Ha, Llion Jones, and Ren Ito. Sakana publicly launched Fugu on June 22, 2026, positioning it not as a single trained foundation model but as a coordinator that dynamically assembles a team of frontier LLMs (GPT-5.5, Gemini 3.1 Pro, and Anthropic's Claude family) to answer each incoming request, presenting the whole system to developers as one OpenAI-compatible endpoint. The launch landed during a two-and-a-half-week window in June 2026 when US export controls temporarily blocked access to Anthropic's Fable 5 and Mythos 5 models, and Sakana marketed Fugu explicitly as a hedge against that kind of single-vendor disruption.
On benchmarks Sakana publishes for the higher Fugu Ultra tier, it scores 73.7% on SWE-Bench Pro, 82.1% on TerminalBench 2.1, 93.2% on LiveCodeBench, 50.0% on Humanity's Last Exam, and 95.5% on GPQA Diamond, describing these as putting it shoulder to shoulder with Fable 5 and Mythos Preview across coding, reasoning, and scientific benchmarks. The base Fugu tier trails Ultra on most axes (59.0% SWE-Bench Pro, 80.2% TerminalBench 2.1, 92.9% LiveCodeBench, 47.2% Humanity's Last Exam), matching Ultra's 95.5% on GPQA Diamond. These are Sakana's own reported numbers; no independent third party had reproduced them at launch, since the models Fugu benchmarks itself against (Fable 5, Mythos Preview) were themselves under export restriction during the comparison window.
Architecturally, Fugu rests on two Sakana papers accepted to ICLR 2026: TRINITY, which splits each incoming task into thinking, drafting, and checking passes handed off between the pooled models, and Conductor, which designs the communication topology between agents and generates each one's instructions. Neither coordinator's own weights are merged into the models it directs. Sakana has not disclosed a parameter count for either coordinator model, nor a context window or maximum output limit for either Fugu tier.
Modality support is text-only as far as Sakana documents it: there is no mention of image, audio, video, or PDF input on the product or release pages. Tool-calling and structured output pass through the same OpenAI-compatible endpoint that handles everything else.
Deployment is direct API only, through console.sakana.ai, plus third-party access via OpenRouter (listed as sakana/fugu-ultra) and integrations with the Vercel AI Gateway, opencode, Creao, and Merge. There is no Bedrock, Vertex, or Azure listing.
Sakana has not published a system card, training data cutoff, or red-teaming disclosure for Fugu itself; safety posture is effectively inherited from whichever underlying model handles a given request, each of which publishes its own safety documentation independently. Weights are fully closed, and Sakana states the routing and orchestration logic is "not exposed by design"; the public GitHub repository (SakanaAI/fugu) contains only a CLI installer script, not any model weights.
Independent reception was more skeptical than Sakana's own framing. A Hacker News discussion following the launch questioned whether Fugu was meaningfully different from OpenRouter's existing multi-model routing, and users reported the $200/month Max plan yielding under three hours of sustained heavy use per week, calling the API noticeably slow and the output quality "nowhere near" the Fable 5 and Mythos Preview models it benchmarks against. Coverage from TechCrunch, VentureBeat, and the-decoder framed Fugu as part of a wider wave of Asian AI labs launching Mythos-adjacent products during Anthropic's export-control gap, describing the reception as cautiously optimistic in the press while noting the harsher grassroots sentiment.
Pricing
Fugu Ultra (pinned build fugu-ultra-20260615) token pricing: $5 per 1M input tokens and $30 per 1M output tokens below 272,000 tokens of context, rising to $10 per 1M input and $45 per 1M output above it; cached input is $0.50 per 1M tokens ($1.00 per 1M above the threshold). Base Fugu bills at the standard rate of whichever underlying model handled the request, with no routing surcharge. Separately, monthly subscriptions cover both tiers: Standard $20, Pro $100 (10x usage), Max $200 (20x usage).
What a real job costs
| Job | Input | Output | Total |
|---|---|---|---|
| Summarise a 20-page PDF | $0.150 | $0.030 | $0.180 |
| Support reply | $0.010 | $0.0090 | $0.019 |
| One coding agent run | $1.00 | $0.600 | $1.60 |
Budgets: 20-page PDF = 30k in / 1k out · Support reply = 2k in / 300 out · Coding agent run = 200k in / 20k out. Computed from the vendor's per-token prices at render time; cached-input discounts are not applied.
Key Features
- Multi-Model Orchestration: Dynamically routes each request across three pooled frontier LLMs using a learned Thinker/Worker/Verifier pipeline, instead of committing to one vendor for every call.
- OpenAI-Compatible Endpoint: Drop-in Chat Completions and Responses API; existing OpenAI SDK clients work with no migration required.
- TRINITY Coordinator: A compact coordinator model, trained with evolutionary optimization, assigns Thinker, Worker, and Verifier roles across the pooled models without merging their weights.
- Conductor Routing: A reinforcement-learning-trained model designs the agent communication topology and generates targeted instructions for each worker model per task.
- Codex CLI Installer: One-line install (curl -fsSL https://sakana.ai/fugu/install | bash) wires Fugu into OpenAI's Codex coding agent under the command codex-fugu.
Pros
- Fugu Ultra's own numbers land close to frontier-tier models on paper, including an 82.1% TerminalBench 2.1 score, though these are all Sakana-reported figures no outside group has reproduced.
- OpenAI-compatible endpoint means zero SDK migration for teams already on Chat Completions or Responses.
- Timed its launch to catch teams stuck without frontier-model access during a vendor disruption, rather than competing head-on in ordinary conditions.
Cons
- Text-only: no documented vision, audio, or video input support.
- No published context window, max output tokens, or system card as of launch.
- Hacker News users reported the priciest subscription tier running out of usage quickly under sustained heavy weekly work, and called the API itself slow.
Benchmarks
- GPQA Diamond: 95.5% vendor-reported · 22 Jun 2026 — PhD-level science questions that are hard to search for, % correct.
- Livecodebench: 93.2 vendor-reported · 22 Jun 2026
- SWE-bench Pro: 73.7% vendor-reported · 22 Jun 2026 — Harder, longer real-repository coding tasks, % solved.
- Terminalbench 2 1: 82.1 vendor-reported · 22 Jun 2026
- Humanity's Last Exam: 50% vendor-reported · 22 Jun 2026 — Expert-written questions across many fields, % correct.
A benchmark is an exam, not the job. Scores transfer unevenly between tasks, so weigh the one closest to your workload and read every figure with its source.
Frequently Asked Questions
What are Fugu's pricing plans in 2026?
Fugu runs on two parallel pricing systems: a monthly subscription that covers both Fugu tiers, or metered pay-as-you-go billing on the higher-accuracy Fugu Ultra tier, where rates step up once a request crosses roughly 272,000 tokens of context. Base Fugu, the non-Ultra tier, bills at whichever underlying vendor model actually handled the request, with no extra routing fee stacked on top. A 500,000-token agentic coding session that crosses that context threshold runs about $9.50 on the metered plan.
Does Fugu have a free plan?
No, Fugu does not offer a free tier or trial credits at launch. The cheapest way in is the metered pay-as-you-go plan, which has no monthly minimum and bills per token used; the $20 Standard subscription is the flat-rate alternative if you would rather not track usage.
What are the best alternatives to Fugu?
The closest comparison is OpenRouter, which offers single-API access to 300+ models from OpenAI, Anthropic, Google and other providers but leaves model selection to the developer instead of routing automatically. Teams that don't need multi-vendor redundancy can also call the underlying vendor APIs directly and skip the coordination layer entirely. Hacker News commenters who tried Fugu argued it doesn't meaningfully improve on what OpenRouter already offers.
Fugu or OpenRouter: which should you pick?
OpenRouter gives you a single API in front of more than 300 models and lets you pick which one handles each call yourself. Fugu instead delegates that choice to its own TRINITY and Conductor models, which assign roles and pick a worker per request without you specifying one. Sakana positions this as a resilience feature against a single vendor going down; Hacker News commenters who tested Fugu argued the automated pick doesn't clearly beat just choosing a model on OpenRouter yourself.
How do you get started with Fugu?
Create an account at console.sakana.ai and generate an API key. Because the endpoint matches the OpenAI Chat Completions format, most teams can start by swapping the base URL and key in their existing SDK client, with no other code changes. If you're setting it up for agentic coding, Sakana also publishes a shell script that adds Fugu as a routing target inside OpenAI's Codex tool, though Windows needs a couple of extra manual steps that Ubuntu and macOS skip. From there, choose either a flat monthly plan or metered token billing depending on how heavy your usage is.
Top Alternatives
- OpenRouter: Pick Fugu if you want automatic vendor failover with zero routing decisions; pick OpenRouter if you want manual control over which model handles each call yourself.
- Claude Fable 5: Choose Fugu for automatic redundancy across vendors; choose Claude Fable 5 directly if you have guaranteed access and want its documented 1M-token context window instead of Fugu's undisclosed one.