GPT-6 Luna India Inference Residency API Setup for Data-Local AI Apps

GPT-6 Luna India inference residency API setup gives eligible organizations a way to route supported model inference to India. OpenAI announced general availability on November 10, 2026, alongside availability in Japan and Australia. The change matters to teams whose application requirements specify where model processing takes place, rather than only where stored data resides. OpenAI release notes

For an application that sends a document image for analysis, the relevant distinction is between storing customer content in a region and processing a supported request there. Inference residency addresses the processing location. An application team still needs to understand which parts of its workflow are covered before treating the feature as an answer to a data-location requirement. Teams defining those requirements may find a project consultation useful for mapping the application workflow.

India residency detail	OpenAI's documented position
Announcement	General availability announced November 10, 2026
Supported interfaces	Responses and Chat Completions APIs
Request modalities	Text and image inputs; text output
Project prerequisites	Approved, active Zero Data Retention; a regional API project; Storage & Processing enabled

What changed for GPT-6 Luna in India

OpenAI's November 10 release notes describe general availability of inference residency for GPT-6 Luna in India, Japan and Australia. For supported requests, processing and model inference take place in the selected country. This is an API capability for eligible organizations, not a blanket location guarantee for every OpenAI workload.

The distinction from data-at-rest residency is operational: storage controls concern where covered content is retained, while inference residency concerns where the model processes a request. A residency requirement that names processing therefore calls for a review of the inference route as well as storage settings. OpenAI's data controls documentation describes these controls at project level.

GPT-6 Luna India inference residency API setup

Start with an API organization that has approved, active Zero Data Retention (ZDR). Configure an India regional project with Storage & Processing enabled, then send its supported requests to the India regional endpoint. Selecting an endpoint by itself does not replace the organization’s eligibility and ZDR approval. OpenAI describes the project controls and regional endpoint in its data controls documentation.

Keep the regional project configuration and the request route aligned. The project setting establishes the relevant residency controls; the endpoint in the application directs the API request to the India route. A request sent through a different route should not be assumed to inherit the intended processing location merely because the application has an India residency requirement.

Send a request through the India endpoint

The following Python example uses the documented regional base URL and the gpt-6-luna model ID with the Responses API. It asks for a short text response and prints the returned text.

from openai import OpenAI

client = OpenAI()
response = client.with_options(
    base_url="https://in.api.openai.com/v1"
).responses.create(
    model="gpt-6-luna",
    input="Reply with OK."
)
print(response.output_text)

In an application, the base URL is part of the request configuration that needs to accompany the correct project credentials. Keep that configuration visible to the team responsible for deployment so a change of environment does not accidentally point the same workload elsewhere.

For teams integrating this route into a broader product, generative AI development can encompass the surrounding application workflow as well as the model call. The routing decision should be recorded alongside the project configuration, so developers can identify which requests are intended to use the India endpoint.

What data-local processing covers

India inference residency concerns where supported model processing happens. It does not mean that every piece of data associated with an application stays in India. To understand the boundary, separate the customer content sent to the API from the system data and infrastructure involved in operating the application.

Customer content includes the material an application submits for a model to process, such as a user’s text or an image. For a supported India regional request, the documented control places model processing in the selected country. Storage at rest is a separate part of the project’s residency configuration, so a requirement covering both storage and processing needs both to be considered.

Trace the application beyond the model call

System data has a different boundary from customer content. An organization should account for that distinction when translating a data-location requirement into an application design; a statement about inference location alone does not establish the location of every category of system data.

The route taken before and after the API request matters too. A user’s device, the application’s own infrastructure and any third-party service are outside the model inference step. For example, an app that accepts a document image may also store the original file or send a result to another service. Those activities need their own location assessment.

For a team planning application development, the practical design question is where each component handles the document and its result. A diagram that shows the device, application backend, model request and any external service can make the boundaries visible to the people responsible for the data-location requirement.

Supported APIs and feature boundaries

OpenAI’s November 10, 2026 release notes say GPT-6 Luna inference residency supports the Responses and Chat Completions APIs. Eligible requests can contain text or image inputs and produce text output. An application can therefore submit an image for analysis, but this regional capability should not be assumed to generate an image in response.

For India regional processing, OpenAI limits GPT-6 Luna to Standard pay-as-you-go processing and does not make hosted tools available. That affects applications that rely on OpenAI-hosted tools as part of a request: their regional workflow must omit those tools or handle the equivalent work within the application.

Plan around the supported request

Consider a document-processing application that accepts an image and returns a text response. The model request fits the listed input and output types. If the application also needs a hosted tool, however, the team must redesign that part of the workflow before treating the request as suitable for India regional processing.

That distinction matters during scope definition. A team can map each planned feature to the API request or to application-managed functionality, then identify dependencies that need a different implementation. Project consultation services can support that planning when the regional requirement affects the wider application design.

Support for this model and its listed APIs does not establish regional support for every model, endpoint, tool or modality. Check the specific combination required by the application before deployment.

Price, rate limits and deployment economics

OpenAI lists GPT-6 Luna at $0.10 per million input tokens and $0.50 per million output tokens, before the regional-processing premium. Its model documentation states that regional processing adds 10% where available. Applying that premium gives effective rates of $0.11 per million input tokens and $0.55 per million output tokens for eligible regional processing.

For a simple illustration, a workload using 10 million input tokens and 2 million output tokens would cost $1.10 for input and $1.10 for output, or $2.20 in total at those regional rates. This example covers model token charges only; it does not estimate the cost of the surrounding application or other services.

Budget for traffic as well as tokens

Token volume determines the model charge in the example, but request timing also matters for capacity planning. OpenAI publishes tier-based limits for requests per minute and tokens per minute, as well as batch limits. A workload with short prompts can encounter a request-rate constraint, while one with longer inputs or outputs can put more pressure on its token allowance. The applicable limits depend on the organization’s usage tier and can change, as can published prices.

Estimate costs using the expected mix of input and output tokens rather than applying one blended rate to all traffic. Output carries a higher listed price per token, so the length of generated responses can materially change the estimate even when the number of requests stays constant. Separate estimates for routine traffic and peak periods also help distinguish a monthly budget question from a capacity question.

During project planning and consultation, these estimates can inform the choice of workload size and the budget allocated to the application around the model. The resulting forecast is most useful when its assumed request volume and token mix are stated explicitly.

Choosing GPT-6 Luna India inference residency API setup for data-local AI apps

Choose India regional processing when in-country inference is required and the workload fits the documented eligibility and feature constraints. As a next step, configure an eligible India project, send a non-sensitive test request through the regional endpoint, and validate the application’s behavior before production.