Join our Newsletter — 33% off our NHI Course
Home› Guides› Sakana AI Models: Every Model, Spec, Price and…
Guide AI Platforms

Sakana AI Models: Every Model, Spec, Price and Release Date

PDF for newsletter subscribers · free to read online

← All guides
By NHI Mgmt Group, under the editorial direction of Lalit Choda Updated 18 November 2026 15 min read
On this page

Sakana AI is an AI research company whose API offers the Sakana Fugu orchestration models, including Fugu Ultra, and the Japanese-specialised Sakana Namazu model. This hub lists every Sakana AI language model named in Sakana's product pages, release posts and official Hugging Face cards as of 17 November 2026, with API IDs, versions, release dates, prices, availability, regional limits and the benchmark figures Sakana reports. It covers Fugu, Fugu Ultra v2.0, Fugu Max v1.0, Fugu Cyber v1.0, Sakana Namazu and the open TinySwallow-1.5B model. It is written for developers, buyers, architects and anyone who needs to know which Sakana model is which. Models are listed by release date, newest first, an objective order that implies no ranking. Every fact comes from Sakana AI's own pages or its official Hugging Face cards, cited in the sources list.

Last checked: 2026-11-17.

This is an independent reference page from NHI Mgmt Group based on public provider documentation; it is not sponsored or endorsed by any provider. It uses Sakana AI's public product documentation as of November 2026. Model details change, so confirm against Sakana AI's current documentation.

Key takeaways

  • Fugu is an orchestration model, not a single model. Sakana describes Fugu as a multi-agent system trained to answer directly or assemble a team of models from a pool, handling selection, delegation, verification and synthesis behind one API.
  • Four Fugu variants are listed. Fugu (default), Fugu Ultra v2.0 (API ID fugu-ultra-v2.0), Fugu Max v1.0 and Fugu Cyber v1.0, plus the single model Sakana Namazu (API ID sakana-namazu).
  • Fugu Ultra costs $5 input and $30 output per million tokens. Rates rise to $10 and $45 above 272K context; Fugu Max is $2 and $6, and Namazu is $0.95 and $4.00 until 1 December 2026.
  • Regional limits apply. Fugu is not offered in the EU or EEA, and Namazu is not offered in the EU, EEA, UK or Switzerland, according to Sakana's product pages.
  • API inputs may be used for training by default. Both product pages say usage data may be used for improvement, with an opt-out in console settings.

What is Sakana AI?

Sakana AI publishes research on model merging, collective intelligence and automated science, and sells products built on it. Its products include Sakana Fugu (multi-model orchestration), Sakana Namazu (a Japanese-specialised LLM with built-in web search and code execution), Sakana Marlin (an autonomous research agent), Sakana Translate (Japanese, English and Chinese translation built on Namazu), Sakana Chat and the Sakana Aquarium agent platform. The API is OpenAI-compatible at https://api.sakana.ai/v1, with keys managed in the Sakana console.

Fugu builds on two ICLR 2026 papers, TRINITY (an evolved coordinator that routes each step of a task to a suitable model) and The Conductor (learning to orchestrate agents in natural language). Facts on this page are as of 17 November 2026.

Current Sakana AI models at a glance

As of 2026-11-17. Source: Sakana Fugu and Namazu product pages and release posts. Models are listed by release date of the current version, newest first.

ModelAPI ID or versionCategoryContextLicenceStatus
Fugu Maxv1.0; API ID not stated by Sakana AI on the pages checked (2026-11-17)Orchestration (cost-performance)Not stated by Sakana AI on the pages checked (2026-11-17)Proprietary serviceAvailable on the API
Fugu Ultrav2.0, fugu-ultra-v2.0Orchestration (maximum performance)Not stated by Sakana AI; pricing tier changes at 272K tokensProprietary serviceAvailable on the API
Sakana Namazusakana-namazuSingle model (Japanese-specialised)Not stated by Sakana AI on the pages checked (2026-11-17)Not stated by Sakana AI on the pages checked (2026-11-17)Available on the API
Fugu Cyberv1.0; API ID not stated by Sakana AI on the pages checked (2026-11-17)Orchestration (cybersecurity reasoning)Not stated by Sakana AI on the pages checked (2026-11-17)Proprietary serviceContact sales
FuguVersion and API ID not stated by Sakana AI on the pages checked (2026-11-17)Orchestration (default, performance and latency)Not stated by Sakana AI; pricing tier changes at 272K tokens from 5 January 2027Proprietary serviceAvailable on the API

"Proprietary service" reflects that Sakana states the models Fugu selects and how it coordinates them are proprietary and not exposed per query; no weights are published for Fugu. See Single Agent vs Multi-Agent Architecture.

Key facts for the main Sakana AI models

As of 2026-11-17. Identical fields for every model; "Not stated by Sakana AI" means not stated by Sakana AI on the pages checked (2026-11-17).

ModelReleasedMax outputKnowledge or training cutoffUnderlying modelsOpen weightsParametersPositioning (provider wording)
Fugu Max v1.02026-09-11Not stated by Sakana AINot stated by Sakana AIFixed pool of open and specialised models, including NVIDIA NemotronNoNot applicable (multi-model system)Optimised for cost-performance across the largest model pool
Fugu Ultra v2.02026-09-11Not stated by Sakana AITraining cutoff 2026-08-28Fixed pool of open and specialised modelsNoNot applicable (multi-model system)Optimised for performance on hard problems
Sakana Namazu2026-08-03 (API launch)Not stated by Sakana AINot stated by Sakana AISingle model tuned from the open Kimi K2.6Not stated by Sakana AINot stated by Sakana AIJapanese-specialised LLM with web search and code execution
Fugu Cyber v1.02026-07-21Not stated by Sakana AINot stated by Sakana AINot stated by Sakana AINoNot applicable (multi-model system)Specialised for cybersecurity reasoning
Fugu2026-06-22 (general availability)Not stated by Sakana AINot stated by Sakana AIDynamic pool; users can exclude models or providers in the consoleNoNot applicable (multi-model system)Balanced for performance and latency; the default for everyday work

How Fugu pools differ by variant

VariantPool typeCustomer control stated
FuguDynamic; the pool can include instances of Fugu itselfExclude specific models or providers in console settings
Fugu UltraFixed pool; Sakana names three frontier models it excludesFixed pool; enterprise customers can request custom configurations
Fugu MaxFixed pool, the largest, routed to the leanest capable modelFixed pool; enterprise customers can request custom configurations
Fugu CyberNot stated by Sakana AIContact sales

Sakana states that per-request token usage and cost are reported, and that because pool agents are swappable, Fugu can route around a provider that restricts access.

Sakana AI model families and naming

  • Fugu: the orchestration family. Variants carry a version (Ultra v2.0, Max v1.0, Cyber v1.0); Fugu Ultra moved from v1.0 at general availability to v1.1 in July 2026 and v2.0 in September 2026.
  • Namazu: a single Japanese-specialised model. Namazu-alpha (March 2026) was a post-training approach for adapting large open foundation models to national contexts; the Sakana Namazu API followed on 3 August 2026.
  • TinySwallow: a small open Japanese model, TinySwallow-1.5B, built with Sakana's TAID distillation method and published in January 2025.
  • Evo models: image and vision-language models built with evolutionary model merging, including EvoSDXL-JP (April 2024), Evo-Ukiyoe and Evo-Nishikie (July 2024) and a Japanese vision-language model (August 2024).
  • Products on top of models: Sakana Translate (upgraded to Namazu on 21 August 2026), Sakana Chat (Fugu and Namazu added on 13 August 2026) and Sakana Marlin (commercial launch 15 June 2026).

When was each Sakana AI model released?

As of 2026-11-17. Dates are those of Sakana AI blog posts.

DateRelease
2026-10-09Sakana Namazu integrated into the Evidence Finder medical search tool
2026-09-11Fugu Max and Fugu Ultra v2
2026-08-21Sakana Translate upgraded to Sakana Namazu
2026-08-13Sakana Chat adds Fugu and Namazu
2026-08-10Gemma 4-based validation of the Fugu orchestrator
2026-08-03Sakana Namazu API
2026-07-24Fugu Ultra v1.1 and Claude Code-compatible endpoints
2026-07-21Fugu Cyber
2026-06-22Sakana Fugu and Fugu Ultra general availability
2026-04-24Sakana Fugu beta
2026-03-24Namazu-alpha
2025-01-30TinySwallow-1.5B (Japanese post; TAID paper submitted 2025-01-28)

How much do Sakana AI models cost?

As of 2026-11-17. Source: Sakana Fugu and Namazu product pages. US dollars per million tokens, pay-as-you-go Token Plan.

ModelInputOutputCached inputScheduled change
FuguUnderlying model's standard rate (top-tier model's rate when several agents run)As input columnAs input columnFrom 5 January 2027 (10:00 JST): $5 / $25 / $0.25 under 272K; $7 / $35 / $0.35 at 272K and above
Fugu Ultra v2.0, under 272K$5.00$30.00$0.50Unchanged on 5 January 2027
Fugu Ultra v2.0, above 272K$10.00$45.00$1.00Unchanged on 5 January 2027
Fugu Max v1.0$2.00$6.00$0.25Unchanged; flat across context lengths
Fugu Cyber v1.0Contact salesContact salesContact salesContact sales
Sakana Namazu$0.95$4.00$0.15From 1 December 2026 (10:00 JST): $1.90 / $6.00 / $0.30

Tool charges and subscriptions

  • Fugu Max tools: web_search and web_fetch cost $0.007 per call.
  • Namazu tools: web search costs $7.00 per 1,000 calls including page contents; code execution costs $0.12 per hour of session time.
  • Fugu subscriptions: Standard $20 a month, Pro $100 a month (10 times the Standard allowance) and Max $200 a month (20 times); each tier includes Fugu, Fugu Ultra and Fugu Max.
  • Currency: Namazu is billed in US dollars with no monthly fee; yen billing is on the Enterprise plan only.
  • Batch pricing: not stated by Sakana AI on the pages checked (2026-11-17).

Orchestration means one request can fan out to several underlying calls, so per-request cost reporting is the control to watch. See the LLMjacking Guide.

Where are Sakana AI models available?

  • Sakana API: OpenAI-compatible endpoint at https://api.sakana.ai/v1, with keys from the Sakana console; existing clients need only a new endpoint and key.
  • Claude Code-compatible endpoints: announced on 24 July 2026 for running Fugu inside a Claude Code terminal workflow; setup is documented in the console.
  • Third-party gateways: the Fugu page lists OpenRouter, Vercel AI Gateway, opencode (models.dev), Creao and Merge; the Namazu page lists OpenRouter and Merge.
  • Apps: Sakana Chat added Fugu and Namazu on 13 August 2026; Sakana Translate runs on Namazu.
  • Regions: Fugu is not available in the EU or EEA, which Sakana attributes to pending GDPR compliance, and is available elsewhere subject to network conditions and local regulation. Namazu is not available in the EU, EEA, UK or Switzerland, and Sakana does not guarantee that data processing stays within Japan.

Restricted-access Sakana AI models

Fugu Cyber v1.0 is priced by contact with sales and is not listed in the Standard, Pro or Max subscription tiers, so access is by arrangement with Sakana AI. Enterprise customers can also request custom model or provider configurations for Fugu. All other variants are listed as available through the API. The regional limits above apply to every Fugu variant and to Namazu.

What is the context window of Sakana AI models?

Sakana AI does not state a maximum context window for Fugu, Fugu Ultra, Fugu Max, Fugu Cyber or Namazu on the pages checked (2026-11-17). Pricing for Fugu Ultra changes at 272K tokens of context, and from 5 January 2027 the same threshold applies to Fugu; Fugu Max is priced flat across context lengths. Because Fugu delegates to underlying models, the effective context depends on the models it selects, which Sakana does not expose per query.

Is any Sakana AI model open source?

Fugu variants are a hosted service, and Sakana states their model selection and coordination are proprietary. Namazu is tuned from the open Kimi K2.6 model, but Sakana does not state a weights release or licence for Namazu on the pages checked. Sakana does publish open models on Hugging Face (49 models in the SakanaAI organisation on the date checked), mostly research artefacts. TinySwallow-1.5B-Instruct carries Apache-2.0 metadata; its card says it is derived from Qwen (Apache 2.0) and trained on Gemma data, so it is also subject to the Gemma Terms and Prohibited Use Policy, and describes it as an experimental prototype for research and development. See the Open-Weight Model Security Guide.

Model cards and system cards

DocumentWhat it states
Sakana Fugu product pageVariants and versions, API ID fugu-ultra-v2.0, pool types, Ultra training cutoff 2026-08-28, prices, regional limits, training-use opt-out, benchmark results
Fugu Max and Ultra v2 release postFugu Max routes to the leanest capable model; switching from earlier Fugu versions is a single-line parameter change; links a technical report on arXiv
Sakana Namazu product pageBase model Kimi K2.6, API ID sakana-namazu, prices and scheduled change, regional limits, training-use opt-out, benchmark notes
TinySwallow-1.5B-Instruct cardTeacher Qwen2.5-32B-Instruct, student Qwen2.5-1.5B-Instruct; TAID distillation then instruction tuning; Japanese; research and development use; Safetensors metadata shows 2B parameters

Sakana AI does not publish a separate system card for Fugu or Namazu on the pages checked (2026-11-17).

Benchmarks

As of 2026-11-17. All figures are provider-reported on Sakana's product pages, where comparison baselines are anonymised as Model A, B and C. No independent leaderboard figures are included.

ModelBenchmarkProvider-reported result
Fugu Ultra v2.0Eight-benchmark suite (including GDP.pdf, Chartography, DeepSWE, Toolathon, SWEFish)Sakana reports the highest or joint-highest score against its anonymised baselines on 5 of 8, and a top-two score on 7 of 8
Fugu Ultra v2.0AutoResearch (bits per byte, lower is better)Mean 0.9774 plus or minus 0.0019; best run 0.9748
Fugu Ultra v2.0Rubik's Cube solver task300 of 300 solved, average 19.72 moves
Fugu Ultra v1.1Gain over v1.0Up to 7.9 points, largest on ProgramBench and Terminal Bench 2.1
Fugu Max v1.0Terminal Bench 2.1, GPQAD, AA-LCR, GDP.pdf, AutomationBench, SWEFishSakana reports the highest overall score on these six against its anonymised baselines, and a cost-performance frontier gain on 7 of 10 benchmarks
Fugu Cyber v1.0CyberGym / CTI-REALM success rate86.9% / 72.1%
Sakana NamazuFairPoliticsQA56.30%, up from 34.10% for the Kimi K2.6 base
Sakana Namazu (customer-reported)119th / 120th Japanese medical licensing exam, excluding image questions96.8% / 96.4% (Aillis Evidence Finder)

Orchestrated results reflect the pool of models available at test time, so they can change when the pool changes. See How to Read LLM Benchmarks.

Deprecations and retirements

No deprecation or retirement dates are published on the pages checked (2026-11-17). The Fugu Max and Ultra v2 release post says existing users switch with a single-line parameter change, and does not say that Fugu Ultra v1.0 or v1.1 is retired. Two scheduled price changes are published: Namazu from 1 December 2026 and Fugu from 5 January 2027, both at 10:00 JST. Pin a versioned ID such as fugu-ultra-v2.0 where one is given, and re-test when Sakana releases a new version.

Data handling and security

  • Training use of API data: Fugu usage data may be used to improve Fugu, and Namazu inputs may be used for training by default; both can be opted out in console settings.
  • Model and provider exclusion: Fugu lets users exclude specific models or providers from its pool, which helps where data must not reach a given provider; Ultra and Max use fixed pools.
  • Opacity per query: Sakana states the models chosen for each query are not exposed, so record the variant and version used instead of the underlying model.
  • Data location: Sakana does not guarantee that Namazu data processing stays within Japan.
  • Retention: not stated by Sakana AI on the pages checked (2026-11-17).
  • Key handling: keep keys out of code, scope and rotate them, and watch per-request cost reports.

See AI Data Retention and Deletion, the Japan APPI Guide and the Data Sovereignty and Residency Guide.

How to choose a Sakana AI model: factual criteria

  • Task type: Fugu is the stated default for everyday work, Ultra for hard multi-step problems, Max for cost-performance and Cyber for cybersecurity reasoning.
  • Language: Namazu is tuned for Japanese and Japanese business tasks.
  • Price: Max is listed at $2 and $6, Ultra at $5 and $30 under 272K, Namazu at $0.95 and $4.00 until 1 December 2026.
  • Pool control: only the default Fugu allows per-customer model or provider exclusion in the console.
  • Region: confirm that your users and data are outside the excluded regions.
  • Single model or system: Namazu runs on its own; Fugu variants coordinate several models.
  • Billing model: pay-as-you-go tokens or a monthly subscription; Cyber is by sales contact.

Practitioner checklist

  • Record the Fugu variant, version and API ID, such as fugu-ultra-v2.0, in the AI inventory.
  • Turn on the training-use opt-out in the Sakana console if your policy requires it.
  • Confirm users and data flows fall outside the EU, EEA, UK and Switzerland limits that apply to each product.
  • Set Fugu model and provider exclusions to match your approved provider list before sending sensitive data.
  • Budget for the 1 December 2026 Namazu and 5 January 2027 Fugu price changes.
  • Monitor per-request token and cost reports, since one orchestrated request can trigger several model calls.
  • Store Sakana API keys in a secrets manager and treat gateway keys the same way. See Model Aggregator and LLM Router Key Security.

Standards and references

Related NHI Mgmt Group resources: Reka Models, Multi-Agent and A2A Security Guide, Foundation Model Selection Guide, LLM Security Guide.

PDF for subscribers

Take this guide with you

Every section, table and checklist in one PDF, with links to the sources and related guides. Subscribe to our weekly NHI & AI Identity Journal and we’ll email you a download link.

Free weekly newsletter

Subscribe to the NHI & AI Identity Journal

The latest on NHI and Agentic AI security – articles, research, breaches, news and events every week.

Bonus 33% off our NHI Course when you subscribe.

PDF Downloads for Newsletter Subscribers

Get the guide

Enter your details and we’ll email you a download link.

The full guide is free to read online. The PDF is a thank you for subscribing.

Fill in all fields marked * and tick both boxes to continue.