Sakana Namazu: Pricing and Setup for the Japanese-Specialised LLM API

What Sakana Namazu is

Sakana Namazu is a Japanese-specialised large language model API that Sakana AI opened for use on August 3, 2026. It is meant to be wired into your own applications and internal systems rather than used through a chat window.

Kimi K2.6 underneath, with extra training for Japanese

The foundation is Kimi K2.6, the open model published by Moonshot AI. An open model is one whose weights — the trained parameters themselves — are published, so anyone can run it in their own environment. Sakana AI layered its own in-house data on top to adapt the model to Japanese and to the way work gets done in Japanese companies.

Accuracy in Japanese was not the only target. The company also lists reducing unnecessary refusals on certain topics and limiting bias in outputs as things it tuned for. In practice, when a model is handling internal documents and email written in Japanese, whether it gets honorific register and business convention right is what decides if it is usable. That is what this tuning is aimed at.

This is not the first time the company has built a product on top of an open model, but the design differs from Sakana Fugu. Fugu orchestrates several frontier models — the top-tier models from multiple vendors — while Namazu is a single additionally-trained in-house model working on its own.

It is the API cut of the model that ran inside Sakana Chat

The name itself is not new. Namazu was already the model inside the company's Sakana Chat service, and this release updates it and exposes it as an API. Sakana AI says it has kept updating the model since Sakana Chat launched, and that this version improves on earlier ones in both the quality of its Japanese responses and its ability to carry out work as an agent — meaning to use tools and keep going without step-by-step instruction.

The company is also open about what prompted the release. Since Sakana Chat launched, it had repeatedly heard the same request: people wanted to call this model from their own products and workflows through an API. Its framing is that the choice of LLM APIs able to handle real work in Japanese remains limited, with frontier models costing a lot and raw open models leaving open questions about quality and data handling. Namazu is pitched at the gap between the two.

Web search and code execution are built in

The other notable piece is that web search and code execution ship as standard tools. You do not have to stand up your own search service or execution sandbox; both are callable from the same API.

The use case Sakana AI leads with assumes exactly that combination: a market research report that lands every Monday morning and runs end to end with no human in the loop, from drawing up the research plan through repeated web searches, cross-checking findings, and writing the report. Chaining several stages while calling tools is the same capability that agent evaluations such as Sakana AI CoffeeBench are built to measure.

Pricing and performance

Billing is per token. A token is the small unit a model uses to read and write text; the volume you send and the volume that comes back become token counts, and the bill scales with them.

Pay-as-you-go with no monthly fee, and tools billed on the side

Sakana Namazu has no monthly plan and no subscription. You pay for the tokens you use, full stop. The company's other service, Sakana Fugu, carries plans from $20 to $200 a month, so the two are billed on different principles entirely.

What deserves attention is that web search and code execution stack on top of the token bill. Sakana AI explains that enabling these tools makes a single API request run several model calls internally. Search results and fetched page bodies are re-sent as input tokens on subsequent internal calls, so input consumption grows. The shared prefix is served from cache at the discounted rate, but the cost structure is still harder to read than plain chat usage. If you are running it as an agent, measure one real workflow before you estimate.

The Japanese benchmarks all came in above the base model

Sakana AI's two claims about performance are that reasoning and problem-solving hold the base model's level, and that Japanese and Japan-specific contexts came in above it. The first is stated to have been checked on AIME26 (mathematical reasoning), MMLU-Pro (broad knowledge and reasoning), and LiveCodeBench v6 (coding).

The standout on the second is FairPoliticsQA, which measures whether answers stay neutral rather than leaning toward the values of any one country. It moved from 34.10% to 56.30%. The company also reports beating the base model on JFBench, which measures instruction following in Japanese, and on a Japanese-English translation task that accounts for Japan-specific proper nouns and honorific language.

These are first-party numbers, and the only comparison is the base model

One caveat before you read anything into the figures. What has been published is Sakana AI's own measurement, and the only point of comparison is the base model, Kimi K2.6. There is no table putting it beside other vendors' Japanese-capable models.

The company notes that the Japanese-English translation task and FairPoliticsQA are internal benchmarks it built itself, which outsiders cannot inspect. JFBench is a public benchmark developed by Preferred Networks, but the measurement came with its own conditions: 200 items sampled from each of the constraint counts 1, 2, 4, and 8. No concrete figures appear in the post for the reasoning benchmarks such as AIME26 — only the claim that the base model's strong performance was maintained.

The direction of travel is as described. But you cannot use these numbers to rank it against another vendor's model, and that is the accurate way to read them.

Setup and constraints

Getting started is deliberately low-friction. There are, however, a few constraints worth checking before you commit.

Rewriting the endpoint and model name is enough

Because it is served as an OpenAI-compatible API, there is no SDK to swap out. What Sakana AI shows is code using OpenAI's Python SDK with the endpoint and key replaced and sakana-namazu given as the model name. The snippet below follows the official sample, with the key read from an environment variable.

import os
from openai import OpenAI

client = OpenAI(
    base_url="https://api.sakana.ai/v1",
    api_key=os.environ["SAKANA_API_KEY"],
)
response = client.chat.completions.create(
    model="sakana-namazu",
    messages=[{"role": "user", "content": "Hello!"}],
)

Calling it from an existing coding assistant follows the same pattern. As with running Sakana Fugu inside Claude Code, swapping the endpoint is the whole trick. Request bodies are assembled as JSON, so tidying the message array and tool definitions visually before you send them makes it much faster to isolate what went wrong when a call misbehaves.

Free ToolJSON Formatter & ValidatorPretty-print or minify JSON data. Catch syntax errors instantly with line numbers and tree view.Try it now →

Not available in the EU, the UK, or Switzerland

Availability is the thing to check first. Sakana Namazu currently cannot be used in the EU/EEA, the UK, or Switzerland. The notice at the top of the product page says the company is still working toward compliance with GDPR and other region-specific regulations.

Where processing happens also needs attention. Sakana AI states plainly that it cannot guarantee data processing is completed entirely within Japan. Workloads that carry a domestic-processing requirement cannot simply be moved onto it. Do not assume that a model built in Japan is one that stays inside Japan.

Inputs feed training unless you opt out

One more setting to check before business data goes anywhere near it. Left at the defaults, your inputs may be used to train and improve the models. Sakana AI says you can opt out at any time from the Console settings, but the default sits on the "may be used" side.

Billing has a constraint too. Charges are denominated in US dollars only, with yen-denominated payment available on the Enterprise plan. If your accounting process has to handle a currency conversion, that is worth squaring with internal rules up front.

Conclusion: where Sakana Namazu fits

The position Sakana AI stakes out is between frontier models and open models — the former described as costly, the latter as leaving open questions about quality and data handling, with Namazu priced to fill the gap. What it should suit is high-volume work where natural Japanese actually matters: internal documents, customer enquiries, and the like. What the published information does not yet give you is grounds to switch away from another vendor on the basis of measured performance.

The practical move is to start with high-volume, Japanese-heavy processing. Swap the endpoint in existing code, run it against your own data, and measure both quality and real cost. Decide from that, and you never have to lean on the benchmark figures. If you are subject to EU, UK, or Swiss availability or to domestic-processing requirements, check the terms before any of that.

For the lineage of the Kimi models underneath, see our explainer on Kimi K3. Sakana AI's other work is covered in its partnership with NVIDIA. When you are assembling requests to send to the API, straightening out the JSON structure first tends to save time.

Free ToolJSON Formatter & ValidatorPretty-print or minify JSON data. Catch syntax errors instantly with line numbers and tree view.Try it now →