OpenAI API Now Supports In-Country Inference in Japan — GPT-6 Luna Only
機械翻訳 / Machine-translated

機械翻訳 / Machine-translated
@aifriends
AI Friends(https://aifriends.jp)のクロスポスト公式アカウント。AIツールの紹介・使い方・できることを、中学生でもわかるやさしい日本語で届けます。
"We want to use AI, but we're worried about our data being processed on overseas servers." Many companies share this concern. This article breaks down what is possible — and what conditions apply — with the in-country inference for the OpenAI API, which launched on November 11, 2026.
On November 11, OpenAI announced general availability of a feature that allows request processing and model inference (the computation by which the AI generates answers) for GPT-6 Luna via the OpenAI API to take place entirely within Japan.
The announcement was made by an OpenAI Japan representative on X and was also covered by ITmedia AI+.
The scope is limited to GPT-6 Luna only. No other models currently support in-country inference.
Anticipated use cases include document summarization, drafting response proposals, and analytical support. Uptake is expected in sectors with strict data-handling requirements, such as finance, telecommunications, healthcare, and government.
In May 2025, OpenAI announced data residency for Japan — a mechanism for storing data within the country. However, that applied to storage only.
This update goes a step further: processing and inference can now also be designated to take place in Japan. Both the storage location and the processing location can now be kept domestic.
Consider, for example, a hospital administrator who wants to use AI to summarize clinical documents. Even if data is stored domestically, processing that occurs on overseas servers may be flagged during internal compliance review. When processing is also domestic, the barrier to justification is significantly lower.
It is worth noting that similar functionality has already been available in Australia and India; Japan is now joining that list.
The only supported model is GPT-6 Luna. It accepts text and image input and produces text output.
Two APIs are available: the Responses API and the Chat Completions API. The only service tier is Standard pay-as-you-go pricing.
According to ITmedia AI+, the rates are $0.10 per million tokens for input and $0.50 per million tokens for output (a token being the unit by which AI reads and writes text). Luna is considered a lightweight model within OpenAI's lineup, making it relatively cost-efficient.
However, according to Xenospectrum, for data residency-compatible models released on or after March 5, 2026, use of regional endpoints incurs a 10% surcharge. Cost estimates should account for this additional charge.
This feature is not immediately available to everyone. According to Impress Watch, the following four conditions must be met:
According to Xenospectrum, the endpoint is jp.api.openai.com. If images are used as input, approval for enhanced ZDR or enhanced MAM is required in addition to standard ZDR.
Note that projects with only MAM (Modified Abuse Monitoring) enabled cannot be used for in-country inference. ITmedia AI+ mentions both ZDR and MAM, but Xenospectrum's breakdown identifies ZDR as the mandatory requirement. It is advisable to confirm with OpenAI directly before applying.
It is easy to assume that "in-country inference" means everything happens within Japan. In practice, there are numerous exclusions:
Impress Watch notes that this does not guarantee all processing will be completed within Japan.
Furthermore, enabling ZDR does not mean that no data is retained at all. According to Xenospectrum, Conversations, Files, Vector Stores, and Batch are excluded from ZDR. In background mode, response data is stored for approximately 10 minutes. Prompt cache may also persist for up to 24 hours.
Consider a scenario where an internal systems administrator is building an AI tool that searches and summarizes internal documents.
The original documents and search indexes would be kept in the company's own environment. Only the relevant excerpts would be sent to GPT-6 Luna. If conversation history is also managed on the company side, the scope of data retained by OpenAI is minimized. Xenospectrum cites this kind of architecture as a recommended design pattern.
For companies in finance or healthcare, being able to state "processing is also domestic" in internal compliance documentation is a significant step forward. On the other hand, since the ChatGPT application is not covered, day-to-day chat interactions used by employees will not benefit from domestic processing.
A good starting point is to pilot GPT-6 Luna on a small scale and verify whether the data retention rules align with internal policies.
This update moves the OpenAI API from "domestic storage only" to "domestic processing as well." However, the limitation of only one supported model remains.
If you need to use a model other than Luna, in-country inference is not currently available. It will be necessary to compare options, including Japan-region offerings from cloud providers. This article has not verified the current state of offerings from other vendors; please consult each provider's official documentation for details on supported models and data handling.
When making comparisons, the following three points provide a useful framework:
A. As of November 12, 2026, GPT-6 Luna only.
A. No. Prior approval from OpenAI for ZDR is required. After that, you must create a new API project with Japan designated and connect via the Japan-specific endpoint.
A. No. ChatGPT and Codex applications are reported to be excluded.
A. This is not guaranteed. System data such as account information and billing data may be stored outside Japan.
A. According to Xenospectrum, use of regional endpoints carries a 10% surcharge.
Start by confirming whether your organization's data falls within the scope of ZDR's conditions and exceptions.
This article is a cross-post from AI Friends.