As of October 21, 2026, OpenAI's models run on all three major US clouds: Azure, AWS Bedrock, and now Google Cloud Model Garden. That turns a model question into an architecture question - the flagship model costs about the same everywhere, so you are no longer choosing a model. You are choosing a billing relationship, an agent framework, a governance posture, and a lock-in profile.
This is a teardown of tools we actually build on. No vendor is "best." Each is best for a specific footprint. Here is how the three differ, and how to pick without painting yourself into a corner.
What just changed
Until this month, running OpenAI models at enterprise scale mostly meant Azure, OpenAI's long-standing primary cloud. That is over. AWS added native OpenAI hosting on Bedrock earlier in 2026, and on October 21, OpenAI and Google Cloud announced that OpenAI's frontier lineup - GPT-6 Astra (heavy research and coding), GPT-6.1 Sol (high performance at lower cost), and GPT-6 Luna (fast everyday tasks) - plus the Codex coding agent, are coming to Google Cloud's Model Garden, billed through the Google Cloud Marketplace against existing GCP commitments. Microsoft remains OpenAI's primary partner and still gets new products first, but exclusivity is gone. For buyers, that means real choice - and a decision to make.
The three platforms, side by side
Azure AI Foundry is OpenAI's premier first-party home. It carries the most complete lineup - the GPT-5 and GPT-6 families, the o-series reasoning models, Sora, and the Realtime APIs - usually first. Its agent framework is built directly on OpenAI's own Assistants API, with fully managed stateful runtimes, Azure AI Search for grounding, and built-in File Search and Code Interpreter. Billing is pay-as-you-go or Provisioned Throughput Units (PTUs) you can carry across models. If your organization is already an Azure and Microsoft 365 shop, this is the path of least resistance and the broadest model menu.
AWS Bedrock is the multi-model broker. It hosts OpenAI's GPT-5 and GPT-6 families alongside Anthropic Claude, Meta Llama, Mistral, and Amazon Nova, and exposes a native OpenAI-compatible endpoint so existing chat-completions code mostly just works. Its agent stack - Bedrock managed agents on the AgentCore Runtime, built on an AWS-native version of OpenAI's Agents API - can run self-hosted or fully managed, with native Knowledge Bases for retrieval. Pricing matches OpenAI's first-party token rates and draws down existing AWS commitments; a committed-baseline tier cuts AgentCore on-demand compute by roughly 22 percent, and components like Knowledge Base vectors, Guardrail evaluations, and the underlying AWS calls are metered separately. If you want one plane to mix OpenAI with Claude, Llama, and Nova in the same agent, this is it.
Google Cloud Model Garden is the newcomer, and its pitch is governance plus Google-native integration. OpenAI's frontier models and Codex land here in private preview, billed through the GCP Marketplace against your existing Google commitments, under Google Cloud's data-governance and access controls - aimed squarely at regulated buyers. The draw is combining OpenAI's models with Google's own (Gemini for business, the Gemini Agent framework, and Google Antigravity) inside one governed environment. If you are already on GCP and need OpenAI capability without leaving your compliance perimeter, this removes a reason to run a second cloud.
| Azure AI Foundry | AWS Bedrock | Google Cloud Model Garden | |
|---|---|---|---|
| OpenAI lineup | Most complete, first (GPT-5/6, o-series, Sora, Realtime) | GPT-5/6 families, native OpenAI-compatible endpoint | GPT-6 Astra/Sol/Luna + Codex (private preview) |
| Other models | OpenAI-first | Claude, Llama, Mistral, Nova (true multi-model) | Gemini + OpenAI, governed together |
| Agent framework | OpenAI Assistants API, managed stateful runtime | AgentCore Runtime (AWS-native Agents API), self or managed | Gemini Agent framework + Antigravity |
| Billing | PAYG or PTUs (portable across models) | First-party token rates, draws down AWS commits | GCP Marketplace, draws down GCP commits |
| Edge it wins on | Broadest, earliest OpenAI access | Multi-model agents in one plane | Governance + Google-native stack |
Price is mostly a wash - the agent layer is not
For raw inference, the flagship tiers are near-identical across platforms. GPT-6 Astra runs about $10 per million input tokens and $50 per million output; GPT-6 Luna about $0.10 input and $0.50 output - the same on Bedrock and Azure. Where real money moves is the agent infrastructure around the tokens: AgentCore session compute and separate metering for retrieval and guardrails on AWS; PTU capacity and per-tool surcharges (File Search and Code Interpreter are billed by the gigabyte per assistant per day) on Azure. Two systems serving identical traffic can land at very different bills depending on how their agent runtimes, retrieval, and tool calls are metered - not on the model's sticker price. Model your workload against each provider's live pricing page before you commit; treat token rates as a wash and scrutinize the agent layer.
How to actually choose
- You are already deep in one cloud. Use it. The draw-down against existing commitments and the governance you already run usually outweigh small cross-cloud differences. Azure shop, stay on Azure; AWS shop, Bedrock; GCP shop, Model Garden now gives you OpenAI without a second cloud.
- You are building multi-model agents. Bedrock is the natural home - OpenAI, Claude, Llama, and Nova behind one endpoint lets you route per task without re-plumbing.
- You sell into regulated buyers. Governance is the deciding factor. Azure and Google Cloud both lead with compliance posture; pick the one your customers already trust and your auditors already know.
- You want the newest OpenAI model the day it ships. Azure still tends to get it first.
Now that the same frontier model runs on every cloud, the model is no longer your differentiator or your lock-in. The agent framework, the retrieval layer, and the tool contracts are.
The move that outlasts the decision
Here is the part we care about most. The teams that stay flexible are the ones who keep their agent logic, prompts, evals, and tool interfaces vendor-neutral, and treat the cloud's agent runtime as a swappable backend rather than the foundation. Hard-wire your product to one provider's Assistants API or one proprietary agent runtime, and you have traded a model lock-in you no longer have for a platform lock-in you chose.
That is how we build: agent architectures that can run GPT-6 on Azure today and route to an open-weight model on your own infrastructure tomorrow, without a rewrite. It is the same lesson this whole market keeps teaching - the model is the easy part; the durable value is the portable, governed integration around it. We made the same case in our teardown of the coding agents, which no longer compete on model quality either.
If you are deciding where to run frontier models, or want your agent stack built so the decision stays reversible, tell us the workflow - we will give you a straight read.