Sakana Namazu: The Complete Guide to Sakana’s Japanese LLM

Sakana Namazu is the other Sakana model — not part of the Fugu family, not an orchestrator, and easy to misunderstand because of both. It’s a Japanese-specialized LLM API with built-in web search and code execution, priced at a fraction of frontier rates, and it happens to be the model most visitors to Sakana’s chat site actually talk to. This is the complete guide: what it is, what it costs (including the tool billing nobody reads), what it’s for, and where it fits alongside Fugu. Facts verified against Sakana’s official pages on August 19, 2026.

What Sakana Namazu is

Namazu (Japanese for “catfish”) began as the model inside Sakana Chat and graduated to a standalone API on August 3, 2026. Sakana’s positioning is precise: a LLM specialized for Japanese language and Japanese business practices, built by taking an open model — from Moonshot AI’s Kimi family — and post-training it on Sakana’s own data for Japanese linguistic and business-context fit, along with tuning to reduce topic avoidance and output bias.

That “built on an open base” fact explains Namazu’s character: it inherits strong general reasoning and agentic capability from its base while adding the Japanese specialization Sakana cares about — and Sakana openly credits the open-model ecosystem it builds on. Sakana pitches it as the middle option in a real gap: frontier models are expensive for Japanese business workloads, raw open models leave quality and data-handling questions; Namazu aims between, at what Sakana calls an accessible price with an OpenAI-compatible API.

Capabilities on the official record: text, image input, file input (PDF, XLSX, CSV, DOCX and more), built-in tools (web search and code execution), extended thinking, and OpenAI-compatible function calling. Context window: 262,144 tokens with 65,536 max output (per OpenRouter’s listing of the model). A new generation shipped on August 13, 2026 — the version now in Sakana Chat — which Sakana says beats the launch version on both Japanese response quality and agentic execution.

Namazu is not Fugu — and when that matters

The two products share a console, a key, and a chat site, which breeds confusion this site spends real effort untangling. Fugu is an orchestrator: it assembles teams from a pool of frontier models and bills orchestration usage. Namazu is a single model: one network, no internal agents, no orchestration line on the bill. Anonymous visitors to chat.sakana.ai get Namazu, not Fugu — so if you evaluated “Sakana’s AI” without logging in, you evaluated Namazu (details here).

Choose Namazu when the work is Japanese-language, high-volume, or tool-driven (search + compute) at low unit cost. Choose Fugu when the work is hard multi-step reasoning worth a team of frontier models. They’re priced an order of magnitude apart, which makes the decision less agonizing than it sounds.

Pricing — including the tools

Per million tokens, as of August 2026: input $0.95, output $4.00, cached input $0.15 — with thinking tokens billed at the output rate. Pay-as-you-go only: no subscription plan includes Namazu.

The built-in tools bill separately, and their accounting deserves attention:

• Web search: $7.00 per 1,000 calls — and each search query and each page-body fetch counts as one call. One “search” that reads five pages = six calls.
• Code execution: $0.12 per hour of retained session.
• The token echo: with tools in play, each internal agent-loop call re-sends the full prompt and tool results so far, billed as input — typically softened by the cache, since the shared prefix bills at the $0.15 cached rate. Tool-heavy requests cost more than their visible tokens suggest, for the same structural reason Fugu’s do.

Model a workload in our cost calculator — Namazu is one of its two models, with the orchestration slider correctly disabled for it.

Quickstart

Namazu uses the same API key as Fugu and the same OpenAI-compatible surface — existing code needs only a base_url and key change. Sakana’s own connectivity check:

export SAKANA_API_KEY={your api key}
curl -X POST https://api.sakana.ai/v1/chat/completions 
  -H "Content-Type: application/json" 
  -H "Authorization: Bearer $SAKANA_API_KEY" 
  -d '{"model":"sakana-namazu","messages":[{"role":"user","content":"こんにちは！"}]}'

The literal model IDs are sakana-namazu (alias) and sakana-namazu-v1.0. On OpenRouter, it’s sakana/sakana-namazu — note the doubled “sakana” — at the same per-token rates as the direct API (verified August 2026), which is exactly what our browser playground uses; Namazu is its cheap second model, labeled there as the pipeline-testing option. The standard Sakana account rules apply (supported regions exclude the EU/EEA, UK, and Switzerland; 18+; training on API data by default with an opt-out).

Reported benchmarks

Sakana’s launch post (August 3, 2026) reports two claims, and both are refreshingly specific. On general reasoning — evaluated on AIME26, MMLU-Pro, and LiveCodeBench v6 — the tuning preserved the strong performance of the open base model rather than trading it away. On Japanese-specific ability, the tuned model beat its base on JFBench (a Japanese instruction-following benchmark developed by Preferred Networks), on a Japanese–English translation task built around Japan-specific nouns and honorifics, and on a neutrality benchmark called FairPoliticsQA — where Sakana reports the largest gain, from 34.10% to 56.30%. The translation and neutrality benchmarks are Sakana’s own internal constructions; the usual vendor-reported caveats apply to all of it.

What Sakana says people should build with it

The launch post’s three worked use cases sketch the intended envelope. First, autonomous market-research reports: a Monday-morning report produced end to end — research planning, repeated web searches, cross-checking, writing — powered by the agentic base plus the built-in search tool. Second, customer support and order-data analysis in Japanese, where the low unit price is the point: many small AI tasks across a company, consolidated affordably. Third — the demo with the most charm — driving a light show of about a thousand robotic fish: given a theme, the model picks motifs, searches for reference images, and choreographs the school via function calling and image recognition, screenshot by screenshot. Business automation and art demo run on the same loop: see, decide, act.

Where you’ll meet Namazu around this site

Namazu is the anonymous default on Sakana Chat; the budget model in our playground and calculator; the cheap connectivity check in the API guide; and the pipeline stand-in we recommend before pointing real workloads at Fugu Ultra. It earns all four roles the same way: real capability at $0.95 per million input tokens.

Context for the rest of the lineup: the Sakana AI company profile and What is Sakana Fugu? Pricing across everything: the pricing guide.