OpenAI Now Runs GPT-6 Luna Inference in India: What It Means for BFSI and SaaS Apps

For a long time, “data stays in India” and “we use a hosted AI model” were hard to say in the same sentence. Many Indian banks, insurers, hospitals and public-sector teams need customer information to be stored and processed in the country. A chatbot or document assistant that sent every prompt to a model running abroad often failed that test, even when the vendor promised careful handling of the stored copy.

On 10 November 2026, OpenAI announced a change that narrows this gap. Inference residency is now generally available for GPT-6 Luna in India, and also in Japan and Australia. For supported requests, both the processing of the request and the model inference take place in the country you select. In plain words: the question you send, and the work the model does to answer it, can stay in India.

This is useful, but it is not a blanket “all of OpenAI now runs in India” switch. It applies to one model, two APIs, and only to organisations that already have approved, active Zero Data Retention. If you build AI features in Laravel or Node.js — the kind of work we do at DayDecoder, a custom software company in Ahmedabad — you should read the limits before you promise a client that their data will never leave the country.

Storage and processing are not the same thing

It helps to separate two ideas that often get mixed up.

Storage is where a copy of your data sits when it is kept: prompts that are saved, files you upload, logs, and so on. Data residency, in the older and narrower sense, is mostly about this. The file is in a country, so you can point to a region on a map.

Processing, or inference, is the live work. Your app sends a prompt. The model reads it, does the computation, and writes a reply. If that computation happens on a machine in another country, the data has still crossed a border, even if no long-term copy is stored there.

For many regulated teams in India, both parts matter. A policy that says “customer data must be processed in India” is not satisfied by storage alone. The 10 November 2026 change is important because it covers processing as well as storage for the supported Luna requests.

OpenAI already offered data residency options in other places, and it already offered India as a place where some customer content could be stored. What is new is that eligible customers can now have the model’s actual computation for GPT-6 Luna happen in India too.

What was announced, in precise terms

Here is the scope, as of 10 November 2026:

GPT-6 Luna is OpenAI’s efficient model for focused, high-volume work. It is a good fit for tasks such as summarising a document, drafting a reply, classifying a support ticket, or answering a question from your own knowledge base. It is not the only model in the GPT-6 family, and this India option does not automatically extend to the others.

Zero Data Retention, often shortened to ZDR, is a control you must be approved for. In simple terms, it keeps customer content out of OpenAI’s usual abuse-monitoring logs, and it forces some APIs not to store responses even if your code asks them to. That is a stronger setting than Modified Abuse Monitoring, which also limits logging but is not enough on its own for regional processing in India, Japan or Australia.

Hosted tools are extras that run on OpenAI’s side, such as some built-in tool-calling features that the platform runs for you. Those are not part of this India processing option. If your app depends on a hosted tool, you will need a different design, or you will need to run the tool yourself and keep the model call within the supported APIs.

Who should care, and who may still need to wait

This change is aimed at organisations that already wanted to use OpenAI, but could not, because of data-localisation rules. The most obvious group in India is banking, financial services and insurance. The same pressure exists in healthcare, government work, and any SaaS product that holds personal or financial records for Indian customers.

A typical example is an internal assistant that answers questions about policy documents, or a customer-support draft tool that reads a ticket and suggests a reply. Until now, some legal and risk teams blocked that project because the inference step happened abroad. With Luna residency, the same team can at least have a concrete, current option to evaluate.

You may still need to wait, or use a different approach, if any of the following is true:

• Your organisation does not have approved, active Zero Data Retention. This is not a self-serve toggle for every account. You have to be eligible and approved.
• Your product needs a model other than GPT-6 Luna, or it needs audio, image generation, fine-tuning, or other APIs that are not in the supported list.
• Your workflow depends on hosted tools that are unavailable in these regional setups.
• You need a processing mode other than standard pay-as-you-go. Batch, flex or fast modes are not what these three regions support for this option.
• Your lawyers need a written contract, a residencies map, and a clear list of subprocessors. A product announcement is not a substitute for that review. Ask OpenAI for the current terms and confirm them against your own policy.

Also remember the basic rule that has been in place for the API since 1 March 2023: data sent to the OpenAI API is not used to train OpenAI models unless you explicitly opt in. ZDR and residency sit on top of that baseline. They do not replace your own duty to send only the data the task needs.

How to plan the move in a real app

The official setup is short to describe and a little longer to do well. This is a practical order we recommend for a Laravel or Node.js product.

1. Confirm ZDR first. If your organisation is not already approved, start that conversation before you rewrite application code. Without active ZDR, the India regional option is not available.
2. Create a new regional API project for India, and select Storage and Processing. Do not assume an old global project will start running in India if you only change the model name. A new project keeps the in-country work separate and easier to audit.
3. Use a new API key from that project, stored in your secret manager or server environment, not in git. Treat the key as production material from day one.
4. Point only the supported calls at this project: Responses or Chat Completions, with gpt-6-luna, and with text or image input and text output. Leave other model calls on their existing project until those models have a residency option you can use.
5. Turn off features this region does not support. If your code enables hosted tools, a different processing mode, or an endpoint outside the supported list, those requests will not be a valid in-country Luna call.
6. Keep Zero Data Retention behaviour in mind. With ZDR on, the store parameter on Responses and Chat Completions is treated as false even if your code sets it to true. If your app currently relies on stored responses, you will need another way to keep the history you need, in your own database in India.
7. Test with real task types, not only a “hello” prompt. Run your actual summarisation, classification and retrieval-augmented generation (RAG) flows. RAG means the app first fetches passages from your own documents, then asks the model to answer using those passages. The documents can stay in your Indian database or search index. The model call still has to go through the regional project.
8. Record the decision. Write down which project, which region, which model, and which endpoints are approved for in-country work. Give that note to security and legal teams. When staff later add a new AI feature, they should not silently point it at a global project.

A note on application design

Residency is easier to keep if you design for it. Put the OpenAI client behind one module in your Laravel or Node.js app, so every call goes through a single place. That module should know which project and which model to use for in-country work. It should refuse to send a request that falls outside the supported shape, rather than quietly failing over to a global endpoint.

Do not log full prompts and replies to a server outside India if the whole point of the project is to keep that content in the country. Check your error trackers, support inboxes and analytics tools. A “safe” model call can still leak data if the application logs are shipped abroad.

If you serve both Indian and overseas customers, you may want two paths: an India regional path for Indian data, and a different project for other regions. Routing should be based on a clear rule, such as the tenant’s country, not on whichever key is handy in the environment file.

What this does not automatically give you

It is easy to over-claim this news in a sales deck. A careful team should be explicit about the limits.

• It does not move every OpenAI product to India. ChatGPT as a consumer app, other GPT-6 models, and many API endpoints are outside this announcement.
• It does not remove the need for access control, encryption, or your own audit logs. The model can be in India and still see data it should never have been sent.
• It does not make an AI answer legally “correct.” If you use the model in a bank or clinic, a human still needs a defined role, and you still need to test for poor answers.
• It does not replace a contract review. Eligibility, subprocessors and current terms can change. Confirm them at the time you go live.

Used honestly, the announcement is still a real step. It gives Indian product teams a current, official way to run a high-volume OpenAI model with in-country processing, for the kinds of text jobs that appear in chatbots, helpdesks and internal search.

How we can help at DayDecoder

DayDecoder is a custom software company in Ahmedabad. We build AI agents, chatbots and RAG systems on Laravel and Node.js, and we work with clients who have to think about where data is stored and processed. If you want help scoping an in-country Luna setup, or reviewing an existing AI feature against these new limits, use the contact form. We reply within one business day.

FAQ

Is every OpenAI request from India now processed in India?

No. As of 10 November 2026, this option covers GPT-6 Luna on the Responses and Chat Completions APIs, for eligible organisations that set up a regional project with Storage and Processing. Other models and endpoints are not automatically included.

What is Zero Data Retention, and why is it required?

Zero Data Retention is an approved OpenAI control that keeps customer content out of normal abuse-monitoring logs and prevents those APIs from storing responses. Official guidance says regional processing in India, Japan and Australia requires active ZDR. Modified Abuse Monitoring alone is not enough.

Can I use images with this option?

Yes, for input. The supported APIs accept text and image input and return text. This is not an image-generation feature.

We already have an OpenAI project. Can we just switch the model name?

You should create a new regional API project in India and select Storage and Processing, then use a key from that project. Changing the model on a global project is not the documented way to turn this on.

Does this also apply in Japan and Australia?

Yes. The same 10 November 2026 announcement made inference residency generally available for GPT-6 Luna in Japan and Australia, under the same kind of setup.

Will this satisfy every Indian data-localisation rule?

Not automatically. It is a technical option for in-country processing of supported Luna requests. Your legal and compliance teams still need to match it against the rules that apply to your business and to confirm OpenAI’s current terms.