New Namazu in Sakana Translate: What 160 Tests Say

Sakana Translate, the free Japanese, English and Chinese translator from Sakana AI, has run on a new generation of Sakana Namazu since August 21, 2026. Sakana AI backs the switch with an in-house test of 160 Japanese-to-English tasks and reports a win rate above 50% against every system it compared. The announcement leaves part of that claim open, and it also gives the three modes, a character limit and a list of planned features.

Sakana Translate now runs on a new Sakana Namazu model

The switch asks nothing of users, because the languages, the features and the price all stayed as they were. On August 21, 2026, Sakana AI replaced the translation model behind Sakana Translate with the new generation of Sakana Namazu.

It describes Sakana Namazu as a model series built with a strong emphasis on understanding Japanese and the culture and context behind it, and says it recently released this generation via API and on Sakana Chat. It credits a refreshed base model and improved training methods for translations that read more naturally.

What changed and what stayed the same

Item	After the update
Model	New generation of Sakana Namazu
Languages	Japanese, English and Chinese, both directions (unchanged)
Price	Free with an account (unchanged)
Features	Translate, Proofread, Ask (unchanged)
User action	None

Its example is omiyamairi, a newborn's first shrine visit, an expression rooted in a Japanese cultural tradition. Sakana AI says the new model grasps what the expression means and renders it into more natural English than other models. The comparison appears as an image only, so the outputs are not in the text.

Is the Sakana Translate app still free?

Yes, and an account is the only requirement, since Sakana AI opens all three features to everyone at no charge. As of the August 21, 2026 announcement, the price was one of the things the update left unchanged.

A free account is enough to paste in a long Japanese text, polish a draft and ask follow-up questions, so anyone who translates Japanese to English now and then can try it on real text without paying. API access is a separate matter, because it appears only on the list of planned features, and no price for it has been announced.

How good is Sakana Translate from Japanese to English?

The only measurement is the vendor's own, and it reports a win rate above 50% against every system compared, scored on 160 in-house tasks. Sakana AI ran the evaluation with TransEvalnia, a method it developed, which judges which of two translations is better on several perspectives such as appropriateness of word choice and naturalness.

The test set is an in-house set of Japanese-to-English translation tasks, and Sakana Translate was compared head-to-head with other translation models. Its output was judged the better translation more than 50% of the time against every system compared.

From this result Sakana AI draws a conclusion of its own. For text with uniquely Japanese expressions and cultural background, it says, a larger general-purpose model does not necessarily translate better, and quality depends on how deeply a model understands Japanese and the culture behind it.

It adds that the results suggest its post-training method is an effective way to improve translation. That reading comes from the party that built both the model and the method.

What the 160-task test leaves unsaid about other systems

The text gives a floor and nothing more, since no competitor is named and no per-system figure is printed. Read literally, "more than 50%" covers everything from a narrow edge to a clean sweep, and the announcement does not say where any one system falls.

The results chart is an image, so its numbers are not in the text either. Both the evaluation method and the test set belong to Sakana AI.

What the announcement's text states and what it leaves out

Item	In the text
Evaluation method	TransEvalnia, developed by Sakana AI
Task set	In-house, Japanese to English
Format	Head-to-head against other translation models
Result	Better in more than half of the comparisons, against every system
Compared systems	Not named
Per-system percentages	Not given

Two outlets filled the gap with names. The AlphaSignal headline says Namazu comes out ahead of Google Translate and DeepL on Japanese cultural nuance, and the Ledge.ai headline, in Japanese, says Sakana AI's own evaluation of Japanese-to-English translation exceeds Google Translate, DeepL and others. Those names are the outlets' reporting, not lines from the announcement.

Anyone who needs a ranking against Google Translate or DeepL cannot get one from this test. The practical step is to put the same Japanese paragraph through each system and compare the results side by side.

Sakana Translate modes and a roughly 5,000-character limit

Each of the three features covers a different job, and only the first has a stated size cap. From the August 21 update onward, all three run on the new model, and users need no extra steps.

The three modes and what each does

Mode	Job	Detail
Translate	Translates a long pasted text in one go	Real-time output, up to roughly 5,000 characters
Proofread	Polishes text into more natural, appropriate phrasing	Highlighted diffs; naturalness, politeness, tone and register
Ask	Answers follow-up questions on a translation or proofread	Same context; nuances, alternative phrasings, grammar and word choice

Proofread shows each change as a highlighted diff, which makes shifts in politeness and register (how formal the wording sounds) easy to spot. Ask keeps follow-up questions in the same context as the translation or proofread, so a specific rendering can be questioned on the spot.

A document of several pages can pass the character limit of Translate mode, and the announcement gives the figure only as "roughly", so it is approximate. A character count taken before pasting gives the exact length, and a text over the limit can be cut at paragraph breaks and translated in parts.

What is planned: file translation, glossaries and an API

Sakana AI gives no date for any of the additions, so none of them belongs in a decision made today. It says it plans to add file translation, which would let users upload PDFs, Microsoft Office files and more for direct translation.

It is also working toward glossary integration, translation models optimized for specific industries, API access, and enterprise capabilities: SSO (single sign-on), audit logs, and on-premises deployment, meaning the software runs on the customer's own servers. Until file translation arrives, a PDF has to be copied into Translate mode as text.

A Sakana Translate API is only planned. The Sakana Namazu models themselves were released via API, as the announcement says, and that API is covered in Sakana Namazu API: pricing, setup and EU limits.

Summary: what changed, what the test shows, what to check

Sakana Translate is free and now runs on a new Namazu model, but the 160-task result is a floor set by Sakana AI itself. Treating the result as a minimum and reading the output against the original is the safe way to use it. The right next step depends on what needs translating:

• Everyday Japanese-to-English text within the character limit: paste it into Translate mode with a free account, then read the result against the original and use Proofread or Ask on any passage that sounds off.
• A ranking against Google Translate or DeepL: the announcement gives none, so compare the same paragraph in each system.
• PDFs, glossaries or an API: all are planned with no dates, so no work should be built around them yet.