GPT-6 Luna regional inference: access, costs and limits
OpenAI’s current API documentation supports GPT-6 Luna regional processing in Australia, India and Japan. The useful change for developers is where eligible requests run, not a new model or permission to move every workload. This guide checks the documented requirements as of November 12, 2026.
Impress Watch reported Japan availability on November 11. OpenAI’s current documentation confirms the capability across all three countries but does not date that change. We are not assigning a single launch date to the three regions.
Eligibility comes before the endpoint
The data controls guide requires active Zero Data Retention, or ZDR. Modified Abuse Monitoring alone is insufficient. Non-U.S. residency also requires approval for abuse monitoring controls and a Modified Retention amendment.
Use a project in the selected region and the matching endpoint: au.api.openai.com for Australia, in.api.openai.com for India or jp.api.openai.com for Japan. Supported processing is limited to gpt-6-luna on Responses and Chat Completions, with text or image input and text output. Only Standard pay-as-you-go is supported. Hosted tools are unavailable, and image support requires enhanced retention-control approval.
For an existing integration, changing a hostname is therefore the last part of the review, not the first. Inventory the actual model, API, service tier and tool dependencies. A workflow that relies on hosted search cannot be treated as unchanged merely because its final answer is text.
Budget for the regional surcharge
OpenAI’s published pricing lists Standard GPT-6 Luna at $0.10 per million input tokens and $0.50 per million output tokens. Regional processing endpoints add 10% for eligible models released on or after March 5, 2026.
Applying that uplift gives $0.11 per million uncached input tokens and $0.55 per million output tokens. A hypothetical request volume totaling one million uncached input tokens and one million output tokens would therefore cost $0.66 for those tokens, before taxes or other applicable charges. This is arithmetic from public list rates, not a customer quote or measured workload.
Estimate the real input-to-output ratio before forecasting a monthly bill. A short classification result and a lengthy generated report can use similar input material while producing very different output volumes. Keep the regional surcharge visible in the estimate rather than hiding it inside an assumed discount.
Regional processing is not an all-data guarantee
The API guide excludes system data and third-party services from residency coverage. Customer infrastructure and end-user location can also put transmission outside the selected region. Storage support alone does not establish local inference.
Draw the full path: source application, request gateway, model call, tool services, logs and downstream storage. Mark which component each supplier’s commitment actually covers. A regionally configured model endpoint does not by itself establish that the entire application meets a legal or internal requirement. Have the responsible team review the complete design.
Do not transfer API rules to ChatGPT
The separate ChatGPT residency guide currently lists inference residency in the U.S., Europe and UAE. For workspaces purchased through AWS Marketplace, it lists U.S. inference only. That purchasing distinction is covered in our ChatGPT Enterprise marketplace guide.
Before migration, confirm eligibility and the intended configuration with the account team, then test a representative non-sensitive workload. Check errors, quality, latency and the records your own application retains. ByteForward has not benchmarked these regional endpoints. Treat the documented availability as a deployment option to evaluate, not proof that a particular migration is ready.
