Key takeaways
- On 21 October 2026, OpenAI and Google Cloud announced that GPT-6 models are coming to Google Cloud's Model Garden. GPT-6 Astra, GPT-6.1 Sol and GPT-6 Luna are in private preview for selected enterprise customers, with general availability "in the coming weeks" (Digital Applied, FinanceFeeds).
- This makes GPT-6 available on all three big clouds. Azure (Microsoft Foundry) and AWS (Amazon Bedrock) both went GA in September (Azure blog, AWS).
- The "global" price is the same on Azure and Bedrock. Astra costs $10 per million input tokens and $50 per million output tokens. Pinning traffic to a specific geography costs about 10% more.
- None of the three offers GPT-6 inference inside India today. From Mumbai on Bedrock, GPT-6 Astra is reachable only through global cross-Region routing, which can send your prompts to any supported region (Bedrock model card).
- Choose by where your data, credits and IAM already live, not by the model. It is the same model on all three clouds. The real differences are billing, governance tooling and data routing.
1. What changed this week
Until September, enterprises that wanted OpenAI's frontier models inside their own cloud account mostly had one option: Azure. AWS brought GPT-6 Astra to Bedrock on 8 September and Sol and Luna on 22 September (AWS what's new). Microsoft made all three GA in Foundry the same month.
Google Cloud was the missing one. Its 21 October announcement, as reported, covers three things:
- Models: GPT-6 Astra (flagship, for coding, computer use and research), GPT-6.1 Sol (general production), and GPT-6 Luna (cheap, high volume). They sit in Model Garden next to Gemini and Claude.
- Billing: you buy through Google Cloud Marketplace. Usage appears on your normal Google Cloud invoice and draws down existing spend commitments. Google's own savings-plan rules mean you don't get a discount on third-party models. The benefit is that the spend counts against a commitment you already made.
- Codex: you can point the Codex CLI and desktop app at Google's platform for inference. AWS added a similar Codex-on-Bedrock option in September.
What's missing so far: there is no public rate card for the Google route, no list of regions, and both zero-data-retention and Confidential Compute are marked "coming soon" (Digital Applied).
2. Price comparison (USD per million tokens, short context)
| Model | Azure Foundry — Global | Azure — US Data Zone | Bedrock — Global CRIS | Bedrock — US Geo / In-Region | Google Cloud |
|---|---|---|---|---|---|
| GPT-6 Astra (in / out) | $10 / $50 | $11 / $55 | $10 / $50 | $11 / $55 | Not yet published |
| GPT-6 Sol (in / out) | $2 / $10 | $2.20 / $11 | See Bedrock pricing | See Bedrock pricing | Not yet published (6.1 Sol in preview) |
| GPT-6 Luna (in / out) | $0.10 / $0.50 | $0.11 / $0.55 | See Bedrock pricing | See Bedrock pricing | Not yet published |
| Cached input (Astra) | $1.00 | $1.10 | $1.00 | $1.10 | — |
Sources: Azure blog and the Bedrock GPT-6 Astra model card. Long-context requests (above 272K input tokens on Bedrock) cost roughly double. Bedrock also offers an "Ultrafast" tier at 6× the standard rate.
The pattern holds across every route: "Global" is the base price, and keeping data inside a named geography costs about 10% more. For Indian buyers this premium doesn't apply yet, because there is no India zone to pay it for.
What it costs in rupees
Take a mid-sized workload: 50 million input tokens and 10 million output tokens a month. That's roughly a support copilot handling a few thousand conversations a day. At Global rates, using an assumed ₹90 per US dollar (check the current rate and add 18% GST on cloud invoices):
| Model | Monthly USD | Approx. monthly ₹ (before GST) |
|---|---|---|
| GPT-6 Astra | $1,000 | ₹90,000 |
| GPT-6 Sol | $200 | ₹18,000 |
| GPT-6 Luna | $10 | ₹900 |
The 100× gap between Astra and Luna matters more than the gap between clouds. Most teams save more by sending routine requests to Luna or Sol than by switching providers.
3. The India data question
This is where most Indian teams should slow down.
- Bedrock, Mumbai (ap-south-1): GPT-6 Astra is listed only as a source region for global cross-Region inference. Requests made from Mumbai can be served anywhere in the global pool. Hyderabad (ap-south-2) isn't listed at all (model card).
- Bedrock, older models: AWS reportedly added an India Geo inference profile in August, routing within Mumbai and Hyderabad. It covers GPT-5.6 Terra and Luna, not GPT-6 (eCorpIT summary). If in-country inference is a hard requirement, that older tier is currently the only OpenAI option on Bedrock.
- Azure Foundry: GPT-6 Standard runs in Global deployments and in US and EU Data Zones. No India Data Zone has been announced (Azure blog, region table).
- OpenAI direct: the
in.api.openai.comendpoint reportedly offers India storage but not India processing. Inference can still happen outside the country (eCorpIT summary). - Google Cloud: no regions published yet.
Under the DPDP Act, 2023, sending data abroad isn't banned by default. However, sector regulators (RBI for payments data, IRDAI, SEBI) and many enterprise customer contracts impose stricter localisation. If you handle payment data or work under a contract that says "data stays in India", none of the GPT-6 routes qualify today. Mask or tokenise personal data before it reaches the model, or use an in-country model for that part of the flow.
4. Which cloud should you call GPT-6 from?
| If you… | Use | Why |
|---|---|---|
| Already run on Azure, use Entra ID, M365 or Purview | Microsoft Foundry | Earliest access to new OpenAI releases under the partnership terms. Purview data policies and Foundry guardrails apply directly |
| Run on AWS with a data lake in S3 and IAM-heavy governance | Amazon Bedrock | Same Global price as Azure. Supports the Responses and Chat Completions APIs, so existing OpenAI SDK code needs only a base-URL change. CloudTrail audits every call |
| Have a Google Cloud commit to burn, or a BigQuery and Gemini stack | Google Cloud Model Garden (once GA) | Spend counts against your commitment, and Gemini, Claude and GPT sit in one catalogue. Wait for the rate card and region list |
| Are a startup with no cloud commitment | OpenAI API direct | Simplest setup and first to get new features. Move to a cloud route when compliance or committed-spend maths demands it |
| Need inference that stays in India | None of the GPT-6 routes yet | Use Bedrock's India Geo profile for GPT-5.6 Terra and Luna, or an open-weights model hosted in an Indian region |
5. What to do this week
- Write down where your inference actually runs. Check whether your current calls use "global" routing. On Bedrock, call
GetInferenceProfile. On Azure, check the deployment type. Put the answer in your data-processing records. - Benchmark Sol and Luna on your real prompts before you default to Astra. At the rates above, sending 80% of traffic to Luna changes your bill far more than switching clouds.
- Keep your code OpenAI-SDK-compatible. Bedrock and Azure both accept the Responses API, so changing provider becomes a config change, and you can move as India regions appear.
- If you're on Google Cloud, fill in the preview interest form but don't plan production around it until the rate card and regions are published.
- Turn on prompt caching. Cached input is 90% cheaper than fresh input on both Azure and Bedrock. That pays off quickly for agents that resend long system prompts.
Bottom line
GPT-6 is now a commodity across the three clouds. The same model, at broadly the same Global price, is available wherever your workloads already run. For Indian teams the real constraint is data locality, not price or availability, and no GPT-6 route solves it yet. Pick the cloud you already govern well, route most traffic to the cheaper tiers, and keep the integration portable so you can switch to an India region as soon as one launches.
Want help choosing, wiring up and governing an LLM stack across Azure, AWS or Google Cloud? See our AI automation development service, or read how to choose an LLM for your product.