Storage was already there. Processing is the new part.
Data residency covers two different promises: where customer content sits at rest, and where the model actually runs when it reads a prompt and writes an answer. OpenAI already offered the first in India. The November 12 announcement is about the second. For supported workloads, the company says both the processing of an API request and the model computation behind the response now take place in India, with supported customer content stored in-country as well.
That distinction is the whole story for the customers this is aimed at. A bank whose compliance review asks where inference happens, not only where files rest, gains something concrete from this release: a direct-API processing-location commitment, added to the India storage residency OpenAI says already existed.
The conditions, in the release's own wording
The qualifying phrase throughout is "supported." The release initially names one model, GPT-6 Luna, and names no second model and no date for one. It also never fully defines which customer content and processing fall inside the residency commitment; selecting Luna is the named condition, not a complete definition.
Supported also means a specific surface. The launch covers direct API usage, through two interfaces: Responses and Chat Completions. ChatGPT and Codex are not mentioned and are not part of the offer as written. The work OpenAI chose to illustrate is text-based: summarizing documents, drafting responses, helping employees analyze information.
And supported means a documented setup that starts over, administratively. Customers get access by creating a new API project with India selected for both storage and processing, then calling the India regional endpoint. The release does not print the endpoint's address, does not define which businesses and developers count as eligible, and does not state whether an existing project can be converted.
What the announcement leaves blank
Three blanks matter more than the headline. The release states no price and no surcharge. Whether regional processing in India costs more than the standard API is not in this text, and this article does not import figures documented for Japan or Europe into the India offer. Customers need the applicable commercial terms before they can compare this route with their existing arrangements.
It also states no retention terms. In-country processing answers where computation happens. It says nothing by itself about how long inputs and outputs are kept, what abuse-monitoring access applies, or which account and system data falls outside the residency promise. Those terms decided the real value of residency offers in other regions, and they will decide this one. Until they are in front of a compliance team, the release provides a location claim, not a complete file.
The release also gives no date for additional models. And there is a comparison it makes against itself. OpenAI notes that a similar residency capability was recently made available on certain models in AWS Bedrock: a separate in-country inference option for OpenAI models in India. This announcement concerns direct OpenAI API access, and it does not establish that the two routes carry identical capabilities or terms. Teams already running the Bedrock route should read this as a second door, not a first one.
Sources
- [1] OpenAI press release, "OpenAI brings local AI processing to India with API inference residency," November 12, 2026, read in full via Business Scanner syndication (identical wording in Business Fortnight and Right Column Media the same day)Read source
- [2] Business Fortnight, same press release, November 12, 2026Read source
- [3] XenoSpectrum, "OpenAI Brings GPT-6 Luna Inference to Japan, With Zero Data Retention Required," November 2026 (the Japan leg of the same expansion, Nov 11; used only to establish that the Japan terms exist separately and are NOT carried into this article's India claims)Read source
- [4] Amazon Web Services, "Introducing OpenAI models on Amazon Bedrock for in-country inferencing in India" (the Bedrock route the release itself references; older GPT-5.6 Terra/Luna profiles)Read source