Key takeaways
- On 12 November 2026 OpenAI launched API inference residency in India. A project pinned to India and called through
in.api.openai.comkeeps both storage and model processing in the country (AIM, OpenAI docs). - Only one model gets in-country processing:
gpt-6-luna. Every other model on the India endpoint gets India storage only, and requests may be processed outside India. - The scope is narrow. Responses and Chat Completions only, Standard pay-as-you-go only, no hosted tools, and Batch and audio endpoints are storage-only. Access goes through OpenAI sales.
- Expect to pay about 10% more. OpenAI charges a 10% uplift on residency endpoints for eligible models released on or after 5 March 2026.
- It is not the only in-country route. Amazon Bedrock's India Geo profile keeps the older GPT-5.6 Terra and Luna inside Mumbai and Hyderabad, and Azure now runs four India regions. Pick by model need, not brand.
1. What OpenAI actually launched
Until this week, India was a storage-only region on OpenAI's API. You could keep stored content in India, but inference could run anywhere. In August that was a common complaint from Indian buyers, and it was reported at the time that only the US, EU and UAE processed locally.
OpenAI's data residency guide now lists India with Processing: Yes, alongside the US, Europe, Australia, Japan and the UAE. For India the guide is specific:
- Endpoint:
in.api.openai.com, used from a project created with India as its region. A project's region cannot be changed later, so this has to be a new project. - Model: in Australia, India and Japan,
gpt-6-lunais the model that supports regional processing. - APIs:
/v1/responsesand/v1/chat/completions, with text and image inputs and text output. Image inputs in these regions need extra approval. - Pricing tier: Standard pay-as-you-go only, so no Flex or Priority processing.
- Access: "Contact sales" is listed for India. Trade press also reports that Zero Data Retention and a contract amendment are needed. Confirm the exact terms with OpenAI before you plan around it.
The official documentation also notes that OpenAI uses Cloudflare Regional Services for TLS termination on the regional hostnames, so that sub-processor stays within the region too.
2. What is not covered
The exclusions matter more than the headline, because most real products use more than one endpoint.
| Feature on the India project | Stored in India | Processed in India |
|---|---|---|
| GPT-6 Luna via Responses / Chat Completions | Yes | Yes |
| Other models (GPT-6 Sol, Astra, older GPT-5.x) | Yes | No |
| Hosted tools (web search, file search, code interpreter) | n/a | Not available for regional processing |
Batch API (/v1/batches) |
Yes | No |
| Audio transcription, translation, speech | Yes | No |
| Flex / Priority processing | Not supported | Not supported |
| System data (usage, billing, analytics) | Excluded from residency | Excluded |
Source: OpenAI data residency guide. The guide also warns that if a region does not support processing for a given request, OpenAI may process and temporarily store content outside it.
The practical consequence: an agent that calls web search, a voice bot that transcribes audio, or a nightly batch job will each still send data out of the country, even on an India project. To get full residency you have to redesign around those gaps, for example by running your own retrieval in an Indian cloud region rather than using hosted file search.
3. Comparing the three ways to get in-country OpenAI models
| OpenAI API (India) | Amazon Bedrock — India Geo | Azure / Microsoft Foundry | |
|---|---|---|---|
| Newest model with India processing | GPT-6 Luna | GPT-5.6 Terra and Luna (reported) | Check the region table per model |
| India locations | Not disclosed | Mumbai + Hyderabad | Pune, Chennai, Mumbai, Hyderabad |
| How you get it | New India project + sales | in. inference profile |
Regional or Data Zone deployment |
| Billing | OpenAI invoice, +10% uplift | AWS bill, existing commitments | Azure bill, existing commitments |
| Hosted tools | Not in-region | Use AWS-native services | Use Azure-native services |
Notes on the other two:
- Bedrock: AWS reportedly added an India Geo cross-Region profile in August for GPT-5.6 Terra and Luna, routing only between Mumbai and Hyderabad (eCorpIT). GPT-6 models on Bedrock are still listed for global routing, as we covered in GPT-6 on AWS, Azure and Google Cloud.
- Azure: Microsoft says the Hyderabad region, India South Central, launched with three availability zones, making four Indian regions (CRN Asia). Foundry's deployment types matter. Global deployments can run in any Azure region, Data Zone stays inside a zone such as APAC (which is wider than India), and only Standard/regional deployments process in the region you pick (Microsoft Learn). Check model availability for your Indian region before you commit.
4. Do you actually need in-country inference?
The DPDP Act does not, by itself, ban processing personal data outside India. It allows cross-border transfer except to countries the government restricts. The pressure comes from sector rules and contracts:
- Payments: RBI's 2018 storage directive requires payment system data to be stored only in India.
- Banking, insurance, capital markets: RBI, IRDAI and SEBI outsourcing and cloud frameworks push regulated entities towards Indian regions and strong audit rights.
- Government and PSU work: MeitY-empanelled clouds and in-country hosting are routine tender conditions.
- Enterprise customers: many large Indian buyers now put "data and processing stay in India" in their security questionnaires, whatever the law says.
If none of these apply, for example a public FAQ bot or code generation on non-sensitive repos, the 10% uplift and the feature limits may not be worth it. If they do apply, this launch removes a real blocker for using a current-generation OpenAI model.
5. What it costs, roughly
OpenAI gives the uplift as a percentage, not as rupee prices, so work it out from your current bill. If a team spends ₹1,00,000 a month on GPT-6 Luna tokens through the global endpoint, the same traffic on the India endpoint would cost about ₹1,10,000 before GST. A cross-cloud data transfer, a redesigned retrieval layer or a dedicated compliance review can easily cost more than that difference in the first quarter. Budget for the engineering work, not just the tokens.
6. What to do this week
- Map your calls. List every OpenAI endpoint, model and hosted tool your product uses. Anything outside "GPT-6 Luna + Responses/Chat Completions" will not be processed in India.
- Classify the data. Separate payment, KYC, health and customer-contract data from public or synthetic data. Only the first group needs the India route.
- Test Luna on your evals. Moving from GPT-6 Sol or Astra down to Luna for residency may cost you quality. Run your own test set before you switch.
- Replace hosted tools. Move retrieval to a vector store in an Indian region (Mumbai/Hyderabad on AWS, Indian Azure regions, or Delhi/Mumbai on Google Cloud) and pass results in as context.
- Start the paperwork early. Sales-gated access and contract amendments take weeks. Ask OpenAI for written confirmation of retention terms and sub-processors for your DPIA.
- Keep a fallback. If you are already on AWS or Azure, compare Bedrock India Geo or an India-region Azure deployment. Your existing cloud commitment discounts may outweigh model freshness.
Need help designing an AI stack that keeps customer data in India? See our AI automation and development service, or our data center and cloud team for India-region architecture.