Model Release·Agentic
Sakana Namazu: Japanese-Specialized API on Kimi K2.6
Sakana Namazu is a closed Japanese-specialized API on Kimi K2.6: OpenAI-compatible, web_search + code_interpreter, $0.95/$0.15 cached/$4.00 per 1M. Not available in EU/EEA, UK, or Switzerland.
At a Glance
- Org: Sakana AI
- Base: Moonshot Kimi K2.6 + Sakana post-train
- IDs:
sakana-namazu(product); console pricing listssakana-namazu-v1.0 - Endpoint:
https://api.sakana.ai/v1(OpenAI-compatible) - Modalities: Text and image in (product pricing covers image tokens); text out; file/tool loops via agent use
- Tools: Built-in web search + code execution (billed separately)
- Context: 256K (catalog packet; confirm in console docs if your stack needs a hard cap)
- Region gate: EU/EEA / UK / CH unavailable for now
Pricing (console, paygo)
| Item | Price |
|---|---|
| Input | $0.95 / 1M |
| Cached input | $0.15 / 1M |
| Output | $4.00 / 1M |
| Thinking tokens | Same as output |
| Web search | $7.00 / 1,000 calls (queries + page fetches each count) |
| Code execution | $0.12 / hour (session retained) |
No monthly plan for Namazu. Agent loops re-send history + tool results as input (cache helps). Source: https://console.sakana.ai/pricing
What's New vs Base Kimi K2.6
- JP business / keigo / customs tuning from Sakana in-house data.
- Built-in search + code tools on the API (not DIY tool wiring only).
- FairPoliticsQA jump 34.10% → 56.30%; beats base on JFBench + internal EN↔JP translation (no absolute scores published for those two).
Caveats
- Closed API; not a drop-in open-weight release of K2.6.
- Default may allow training on inputs; opt out in Console (vendor FAQ).
- Data residency in Japan is not guaranteed (vendor FAQ).
- EU/EEA/UK/CH blocked until compliance work lands.
- Retention of AIME26 / MMLU-Pro / LiveCodeBench v6 is vendor claim without published Namazu absolute numbers on the announce page.
Sources
Questions
Frequently Asked Questions
- A closed API from Sakana AI (Aug 3, 2026 per catalog): Japanese-specialized post-training on Kimi K2.6, with built-in
web_searchandcode_interpreter, OpenAI-compatible athttps://api.sakana.ai/v1. Product IDsakana-namazu; console pricing also listssakana-namazu-v1.0. - Paygo: $0.95 input / $0.15 cached input / $4.00 output per 1M tokens. Thinking tokens bill at output rates. Web search is $7.00 / 1,000 calls (queries + page fetches each count). Code execution is $0.12 / hour (session retained). No monthly plan for Namazu. Source: console.sakana.ai/pricing.
- Not available in the EU/EEA, UK, or Switzerland pending GDPR work. Region gate applies until compliance work lands.
- 256K per the catalog packet; confirm in console docs if your stack needs a hard cap.
- The vendor says it keeps K2.6-class AIME26 / MMLU-Pro / LiveCodeBench v6 while lifting JP benches (FairPoliticsQA 34.10% → 56.30%). Absolute scores for JFBench and internal translation are not published. Retention claims for the English benches lack published Namazu absolute numbers on the announce page.
- Default may allow training on inputs; opt out in Console (vendor FAQ). Data residency in Japan is not guaranteed (vendor FAQ).
Continue Reading