by Sakana AI

Fugu review, pricing and limits

Sakana AI's multi-agent orchestration API, sold as four models behind one OpenAI-compatible endpoint, for teams that want frontier-level output without depending on a single model vendor.

Best for
Hedging against a single vendor being unavailable or export-restricted
Not for
Vision or audio applications
  • Available
  • Proprietary

Should you use Fugu?

Pick Fugu if you want one OpenAI-compatible endpoint that keeps working when a single AI vendor is cut off, or the $2/$6 Fugu Max pool for bulk agent runs. Skip it if you are in the EU or EEA, where none of the four models is offered, or if you need a published context window and system card.

Best for

  • Hedging against a single vendor being unavailable or export-restricted Launched in June 2026 during a window when US export controls blocked two Anthropic models; Fugu Ultra v2.0's pool avoids depending on any single proprietary frontier model.
  • Agentic coding workloads The June build scored 73.7% on SWE-Bench Pro in Sakana's technical report, and a Codex CLI installer ships with it.
  • Cost-sensitive bulk agent runs Fugu Max v1.0 is $2 per 1M input and $6 per 1M output tokens at any context length.

Not for

  • Vision or audio applications Fugu is documented as text-only; no image, audio, or video input is listed.
  • Cost-sensitive high-volume production Hacker News users reported the $200/month Max plan capping under 3 hours of heavy weekly use and called the API slow.
  • Teams needing a guaranteed fixed model or SLA Fugu routes each request across a pool of models dynamically; you cannot pin one specific underlying model per call.

Fugu pricing

Input, per 1M tokens
$5.00/1M
Output, per 1M tokens
$30.00/1M
Cached input, per 1M tokens
$0.50/1M
Pricing model
subscription + per-token
Free tier
No

What a real job costs

JobInputOutputTotal
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.

Cost examples

Code review of a 150K-token repo diff (Fugu Ultra, under 272K)
$0.75
Agentic coding session, 500K tokens in / 100K out (Fugu Ultra, above 272K)
$9.50
One month of light daily use
$20.00

Read on Sakana's pricing pages on 9 November 2026. Subscriptions are Standard $20, Pro $100 (10x Standard usage) and Max $200 (20x) a month, and every tier includes Fugu, Fugu Ultra and Fugu Max. Pay-as-you-go: Fugu Ultra is $5 per 1M input and $30 per 1M output tokens below 272K tokens of context ($10 and $45 above), cached input $0.50 ($1.00); Fugu Max is $2 input, $6 output and $0.25 cached at any length, with web search and fetch at $0.007 a call. Base Fugu bills at the rate of the model that handled the request until 5 January 2027, then $5 input and $25 output below 272K ($7 and $35 above); from that date the new rate also draws down subscription allowances. Fugu Cyber is quoted by sales. Pay-as-you-go tokens are served at higher priority than subscription tokens.

What Fugu can do

Sakana Fugu is a multi-agent orchestration API from Sakana AI that routes each request across a pool of models. It comes as four models, Fugu, Fugu Ultra v2.0, Fugu Max v1.0 and Fugu Cyber; Fugu Max costs $2 per million input tokens, and Sakana reported Fugu Ultra at 73.7% on SWE-Bench Pro in June 2026.

What goes in and what comes out

Reads

  • text
  • tool-calls

Writes

  • text
  • tool-calls

Capabilities

  • Function callingSupported
  • Structured outputSupported
  • Tool useSupported
  • Audio inputNot supported
  • Audio outputNot supported
  • Code executionNot supported
  • Video inputNot supported
  • VisionNot supported
  • Web browsingNot supported
Reasoning modes
standard

Key features

  • Four models, one API

    Fugu, Fugu Ultra, Fugu Max and Fugu Cyber share one OpenAI-compatible endpoint, so a caller switches model by changing one parameter.

  • Multi-Model Orchestration

    Each request is split across a pool of models by a learned coordinator; Fugu Max's pool includes NVIDIA's Nemotron family among its open-weight models.

  • Pool opt-out

    On base Fugu you can exclude named providers or models from the pool; Fugu Ultra and Fugu Max keep fixed pools.

  • TRINITY and Conductor

    Two ICLR 2026 methods: an evolved coordinator assigns Thinker, Worker and Verifier roles, and an RL-trained Conductor writes each agent's instructions.

  • Codex CLI Installer

    A one-line install script adds Fugu to OpenAI's Codex coding agent under the codex-fugu command.

Known limitations

  • Not available in the EU or EEA, per Sakana's product page in November 2026.
  • No published context window, maximum output length or system card for any Fugu model.
  • Text and tool calls only; Sakana documents no image, audio or video input.

Who it is for

  • Engineering teams wanting one endpoint that survives a single vendor outage or cutoff
  • Teams running bulk agent workloads who want Fugu Max's lower token price
  • Developers building agentic coding tools through the Codex CLI integration
  • Security teams evaluating Fugu Cyber through Sakana's sales team

Supported languages

  • English

Fugu benchmarks

Standout result

$6

Fugu Max v1.0 output price per 1M tokens

Source: Sakana AI pricing page, read 9 November 2026

Where it sits

Pricier than 87% of the 104 GA models with a published price, and in the top third on GPQA Diamond (rank 3 of 63).

MeasureThis modelRankMedian
Blended price (3:1)$ per 1M tokens$11.25/M#89 / 104peer median $1.73/M
GPQA Diamond% correct95.5%#3 / 63peer median 87.8%

Ranks are against GA models on Hokai that publish the same figure; ties share a rank.

Benchmark scores

  • GPQA Diamond vendor-reported · 09 Nov 202695.5%PhD-level science questions that are hard to search for, % correct.
  • LiveCodeBench vendor-reported · 09 Nov 202693.2
  • Terminal-Bench 2.1 vendor-reported · 09 Nov 202682.1%Multi-step tasks completed in a real command line, % solved.
  • SWE-bench Pro vendor-reported · 09 Nov 202673.7%Harder, longer real-repository coding tasks, % solved.
  • Humanity's Last Exam vendor-reported · 09 Nov 202650%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.

Strengths and limits

Strengths

  • Fugu Max gives a cost-first option at $6 per 1M output tokens without giving up the orchestration layer.
  • OpenAI-compatible endpoint means no SDK migration for teams on Chat Completions or Responses.
  • Designed to keep working when one vendor is rate-limited, cut off or export-restricted, which a single-vendor API cannot do.
  • OpenAI-compatible endpoint, so existing SDK clients switch by changing the base URL and key.
  • Base Fugu lets you remove specific providers or models from its pool for data, privacy or compliance reasons.

Limits

  • Sakana's September builds come with benchmark rankings, not a published score table, so like-for-like comparison is hard.
  • Base Fugu's January 2027 switch to fixed rates can shrink what a subscription covers.
  • Hacker News testers said the top subscription drained within a few hours of heavy weekly work and that responses felt slow.

Fugu alternatives

ModelLabContextIn / out per 1MReads and writes
Fugu This pageSakana AI--$5 / $30text → text
OpenRouterPick 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 5Choose 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.Anthropic1M$10 / $50text, image, pdf → text
Gemini 4 HydrogenGoogle DeepMind's November 2026 Gemini 4 model for coding agents, with four thinking levels from low to xhigh and a free tier in AI Studio.Google1M$1.50 / $7.50text, image, audio, video, pdf → text

Fugu specs

Architecture
Multi-agent orchestration (TRINITY coordinator assigns Thinker/Worker/Verifier roles; Conductor routes agent topology)
Mixture of experts
No
Parameters
Not disclosed (lightweight coordinator model routing across a pool of third-party frontier LLMs)

Training data

Not disclosed by Sakana AI. Fugu's own TRINITY coordinator is trained via evolutionary optimization and its Conductor router via reinforcement learning; the pooled worker models (GPT-5.5, Gemini 3.1 Pro, Claude) are trained independently by their respective vendors, who do not share training data details with Sakana.

API endpoints

NameEndpointAuth
Sakana APIhttps://console.sakana.ai/loginAPI Key
OpenRouterhttps://openrouter.ai/sakana/fugu-ultraAPI Key

SDKs

  • Any OpenAI-compatible SDK (Python, TypeScript, etc.)

Prompting patterns

Show 1 prompting pattern
  • Standard OpenAI Chat Completions format

    { role: 'user', content: '...' } via the OpenAI-compatible endpoint

    Fugu is designed as a drop-in swap for existing OpenAI SDK clients.

Known quirks

Show 5 known quirks
  • Behaves like a premium model router

    Hacker News commenters questioned whether Fugu meaningfully differs from OpenRouter's existing multi-model routing, since it does not train its own frontier-scale worker models.

    Workaround: Test cost-per-task on the pay-as-you-go token plan before committing to a subscription tier, and compare directly against OpenRouter or a single vendor's API for your specific workload.

  • Ultra price steps up above 272K context

    Fugu Ultra v2.0 charges $10 input and $45 output per 1M tokens once a request passes 272K tokens of context, against $5 and $30 below it.

    Workaround: Chunk long documents or repos to stay under 272K tokens per call where the task allows it, to stay on the base rate tier.

  • Max subscription usage caps bite fast

    Independent users reported the $200/month Max plan yielding under 3 hours of sustained heavy agentic use per week.

    Workaround: Track usage against the plan's stated multiplier early in the billing cycle and fall back to the metered token plan if the subscription cap is reached before month's end.

  • Base Fugu billing changes on 5 January 2027

    Base Fugu moves to fixed rates ($5 input, $25 output per 1M below 272K) and the same rate draws down subscription allowances, so a plan may cover less work after that date.

    Workaround: Pin fugu-v1.0 (available from 1 December 2026) if you need today's configuration, and re-check usage against your renewal date before January.

  • Not available in the EU or EEA

    Sakana's product page says Fugu is not offered in the EU or EEA while it works toward GDPR and EU-specific compliance.

    Workaround: European teams should route through a vendor with an EU offering, or confirm with Sakana sales before planning a rollout.

Fugu facts

Deployment
API
Licence
Proprietary (Sakana AI)
Safety posture
Balanced
Released
22 Jun 2026
Current version
fugu-ultra-v2.0
Knowledge cutoff
28 Aug 2026
Family
Fugu
Variant
Chat
Data retention
Not disclosed by Sakana AI for Fugu specifically at launch.

Content restrictions

  • No Sakana-specific content policy published; refusal behavior is inherited from whichever pooled model (GPT-5.5, Gemini 3.1 Pro, Claude) handles a given request.

Safety controls

Rlhf
No
System prompt
Yes
Constitutional AI
No
Content filter API
No
Moderation endpoint
No
Red teaming
Not independently disclosed by Sakana AI; pooled models publish their own safety cards separately.

Data governance

Iso 27001
No
Soc2 type2
No
GDPR compliant
No
HIPAA eligible
No

Version history

Show all 5 versions
  • Fugu public launch · 22 Jun 2026

    Public release of Fugu and Fugu Ultra with an OpenAI-compatible API, subscription and token pricing, and a Codex CLI installer.

  • fugu-ultra-20260615 · 15 Jun 2026

    Pinned Fugu Ultra build referenced in Sakana's launch-day pricing and benchmark tables.

  • fugu-ultra-v2.0 · 11 Sep 2026

    Fugu Ultra v2.0: pool of open and specialist models, Fable 5, Fable 5.1 and GPT-6 Astra excluded; training cutoff 28 August 2026.

  • fugu-max-v1.0 · 11 Sep 2026

    New cost-first model with the largest pool, including NVIDIA Nemotron; $2/$6 per 1M tokens.

  • fugu-v1.0 (announced) · 01 Dec 2026

    Sakana says fugu-v1.0 is added on 1 December 2026 to keep the current base configuration; fugu stays an alias for the latest build, and base Fugu moves to fixed rates on 5 January 2027.

What changed

  • Pricing change Subscription and token pricing published (fugu-ultra-20260615) source
  • Release Fugu and Fugu Ultra publicly launch (Fugu public launch) source

Questions about Fugu

How much does Sakana Fugu cost in November 2026?

Subscriptions run $20, $100 and $200 a month and each one covers Fugu, Fugu Ultra and Fugu Max. Metered use is cheapest on Fugu Max at $2 per million input and $6 per million output tokens, while Fugu Ultra v2.0 charges $5 and $30 below 272K tokens of context. Fugu Cyber is priced by Sakana's sales team.

Is there a free way to try Fugu?

No. Sakana's pricing page lists no free tier or trial credits. The lowest-commitment start is pay-as-you-go with an API key from console.sakana.ai, billed per token with no monthly fee; the $20 Standard plan is the flat-rate option for occasional hands-on use.

What changes in Fugu's pricing on 5 January 2027?

Base Fugu stops billing at the rate of whichever model handled the request and moves to fixed rates: $5 input and $25 output per million tokens below 272K of context. Fugu Ultra, Max and Cyber prices stay the same. Subscription fees do not change, but the new rate applies to allowance use, so the effective allowance may shrink.

Which Fugu model should I pick: Fugu, Ultra or Max?

Fugu is the low-latency default for coding help and chat services, and it is the only one where you can remove providers from the pool. Fugu Ultra v2.0 targets hard multi-step work where quality matters more than cost. Fugu Max v1.0 orchestrates the largest open-weight pool at the lowest token price, which suits bulk agent runs.

Where does OpenRouter beat Fugu?

OpenRouter lets you choose the exact model for every call and works in more places, while Fugu decides the routing itself and is not available in the EU or EEA. If you need a documented context window, a fixed model per request or a European deployment, OpenRouter or a direct vendor API is the safer route.

About Fugu

Fugu is a multi-agent orchestration system from Sakana AI, the Tokyo research lab founded in July 2023 by David Ha, Llion Jones and Ren Ito. Sakana launched it on 22 June 2026 as a coordinator that assembles a team of models for each request behind one OpenAI-compatible endpoint, rather than as a single trained foundation model. On 11 September 2026 it added Fugu Max v1.0 and Fugu Ultra v2.0, and the product page now lists four models: Fugu, Fugu Ultra, Fugu Max and Fugu Cyber.

Read the full background

The four models split by job. Fugu is the low-latency default for everyday coding, code review and chat services. Fugu Ultra v2.0 aims at answer quality on long multi-step work; Sakana says it does not depend on any single proprietary frontier model, that its pool leaves out Claude Fable 5, Fable 5.1 and GPT-6 Astra, and that its training cutoff is 28 August 2026. Fugu Max v1.0 widens the pool with open-weight and specialist models, including NVIDIA's Nemotron family through a collaboration with NVIDIA, and is priced for cost: Sakana compares its output price with Claude Sonnet 5, GPT-5.6 Terra and Kimi K3. Fugu Cyber targets security analysis and vulnerability research, with Sakana reporting 86.9% on CyberGym and 72.1% on CTI-REALM, and is sold through sales only.

The benchmark scores stored on this page are for the June 2026 build (fugu-ultra-20260615), from Table 1 of Sakana's technical report: Fugu Ultra at 73.7% on SWE-Bench Pro, 82.1% on Terminal Bench 2.1, 93.2% on LiveCodeBench, 50.0% on Humanity's Last Exam and 95.5% on GPQA Diamond, against provider-reported baselines for Claude Opus 4.8, Gemini 3.1 Pro and GPT-5.5. For the September builds Sakana publishes rankings rather than a score table: Ultra v2.0 best or joint best on five of eight benchmarks, Max v1.0 best on six, including Terminal Bench 2.1 and GPQA Diamond. All of these are Sakana's own figures.

Architecturally, Fugu rests on two ICLR 2026 papers. TRINITY is a lightweight evolved coordinator that assigns Thinker, Worker and Verifier roles across several turns, and the Conductor is trained with reinforcement learning to design how agents talk to each other and what each is told. On base Fugu you can opt specific providers or models out of the pool for data, privacy or compliance reasons; the Ultra and Max pools are fixed. Sakana has not published a context window, a maximum output length or a system card for any Fugu model, and modality support is text and tool calls only as far as its pages document.

Access is the API at console.sakana.ai, Fugu Ultra on OpenRouter and the Vercel AI Gateway, and a one-line installer that wires Fugu into OpenAI's Codex coding agent. None of the Fugu models is available in the EU or EEA, which Sakana says will last until it completes GDPR and EU-specific compliance work. Pricing changes on 5 January 2027, when base Fugu stops billing per underlying model and switches to fixed token rates; on 1 December 2026 Sakana adds fugu-v1.0 so callers can keep the current configuration, while the plain fugu ID stays an alias for the newest build.

The case against it is mostly about proof and control. A Hacker News discussion after the June launch asked whether Fugu differs meaningfully from manual routing on OpenRouter, and users reported the $200 Max plan running out after under three hours of heavy use a week and the API feeling slow. A team that already has steady access to one frontier vendor and wants a documented context window and a published system card is better served calling that vendor directly; a team that wants one endpoint that keeps working when a vendor is cut off, or the cheaper Max pool for bulk agent work, is the buyer Fugu is built for. The best AI agents guide shows where it sits among coding agents.

Sources and guides

More AI Models on Hokai

  • Qwen4-Max-PreviewAlibaba's November 2026 Qwen 4 flagship preview, open only to Token Plan subscribers, with three reasoning levels from low to xhigh.
  • Claude Fable 5.1Superseded by Claude Fable 5.2 in November 2026 but still sold on the Claude API: Anthropic's second Mythos-class model, released Sept 1, 2026, for long-horizon coding and research.
  • Gemini Decision 1Google DeepMind's decision model scores many questions over one multimodal input of up to 262,144 tokens and bills zero output tokens.
  • Shieldstral 1.0Mistral's open-weight multimodal safety classifier, released 4 August 2026, that screens text and images against policies you write.
  • Claude Mythos 5.2Anthropic's November 2026 Mythos model: Fable 5.2's weights with relaxed cyber and biology safeguards for verified teams, with a 1M-token context.
  • ERNIE-5.5-Flash-PreviewBaidu's text-only preview mixture-of-experts model, served through Qianfan on OpenRouter with a 1M-token context window and tool calling.