OpenAI brings its frontier models and Codex to Google Cloud

What OpenAI and Google Cloud announced

OpenAI's frontier models are coming to Google Cloud. In a post published on 21 October 2026, OpenAI said it is partnering with Google Cloud to bring its frontier models and its Codex coding agent into Google's Model Garden, the catalogue Google Cloud customers already use to pick and deploy AI models. OpenAI models will also come to Google Antigravity, Google's agent-first coding tool, and to the Gemini agent, Google's workplace agent.

General availability is planned "in the coming weeks," according to OpenAI, and companies can request preview access now through a form on OpenAI's site. The post ends with a line aimed at Microsoft's investors as much as anyone: "Microsoft remains OpenAI's primary cloud partner."

The deal means OpenAI's paid models will be sold through all three of the largest US cloud providers. CNBC reported that Amazon began hosting OpenAI models in April, the day after Microsoft's exclusive licence to OpenAI's intellectual property ended, and AI Lookout has covered how Bedrock has kept adding OpenAI features since then. Google is the last of the three to sign on.

How customers will buy it

The commercial details matter more than the press language. OpenAI says Google Cloud customers will buy its frontier models through Model Garden, with the transaction handled by Google Cloud Marketplace. Enterprises with Google Cloud spend commitments will be able to count those purchases against their commitments.

That last point is the one procurement teams will notice. Large companies often sign multi-year cloud contracts in which they promise to spend a set amount. Until now, a business that wanted OpenAI's newest models typically needed a separate contract with OpenAI, or a route through Microsoft Azure. Buying through Google lets a company that has standardised on Google Cloud use OpenAI without a new vendor relationship and while drawing down money it has already committed.

OpenAI pitches the arrangement mainly at regulated industries. Its post says the integration will pair OpenAI's models with Google Cloud's access controls, centralised data governance and infrastructure. It does not say where the models will physically run, whether requests are served from Google's data centres or passed to OpenAI's own, or which Google Cloud regions will be offered. Those details usually decide whether a bank or a hospital can actually use the service.

Antigravity and the Gemini agent

The less expected part of the deal is inside Google's own agent products. Antigravity is Google's agent-first coding environment, which Google DeepMind has used for demonstrations such as a recent antimicrobial peptide search run by an Antigravity agent. According to Antigravity's model documentation, its model menu has so far offered Gemini models plus Anthropic's Claude and OpenAI's open-weight gpt-oss model. Adding OpenAI's paid frontier models puts them next to Gemini in a Google-branded tool.

The same goes for the Gemini agent, which Google Cloud chief executive Thomas Kurian introduced at Gemini at Work on 8 October. Google described that agent as able to route jobs to the best model for each task, and at launch the third-party option it named was Claude. OpenAI says its models will "soon" be available there too, "providing enterprise customers with model choice flexibility."

Codex is the other headline item. OpenAI's coding agent has become one of its main enterprise products, and the post says Codex agentic workflows will be part of the Model Garden integration. OpenAI did not explain how Codex will be packaged on Google Cloud, for example as a hosted agent, as models tuned for coding that Antigravity calls, or both.

Which models, and when

OpenAI's post refers only to its "frontier models" and does not name them. Several outlets described the deal as bringing GPT-6 Astra to Google Cloud, but AI Lookout could not find that model name in OpenAI's announcement. OpenAI's current API line-up is GPT-6 Astra at the top, the cheaper GPT-6.1 Sol for complex coding and computer use, and GPT-6 Luna for high-volume work, so it is reasonable to expect some or all of those, but the exact list has not been confirmed.

Google's own documentation has not caught up yet. On 21 October its OpenAI models page for the Gemini Enterprise Agent Platform still listed only the open-weight gpt-oss 120B and gpt-oss 20B models, which have been available there as managed APIs for some time. Pricing has not been published either. Whether Google Cloud will match OpenAI's own API prices, or charge more for the managed service, is one of the main questions for buyers.

Why OpenAI wants every cloud

For OpenAI, the logic is distribution. Axios noted that the agreement gives OpenAI another way to reach corporate customers without asking them to set up a separate relationship with it. OpenAI is trying to grow its business revenue alongside ChatGPT's consumer base, and CNBC reported that it is courting investors at a $1.4 trillion valuation ahead of a planned stock market listing. Enterprise customers who already spend heavily on Google Cloud were, until now, easier for Google and Anthropic to reach than for OpenAI.

The two companies are not strangers. Quartz and CNBC both recall that OpenAI said last year it was using Google's cloud infrastructure to run ChatGPT and its API. That was OpenAI buying computing power from Google. This deal runs the other way: Google selling OpenAI's models to its own customers.

What Google gets out of selling a rival

Google is now in the odd position of reselling a direct competitor to Gemini. Kurian's quote in OpenAI's post frames that as customer choice: Google wants to offer "a broad selection of models, including those from OpenAI, so they can choose the right model for any given job." Google Cloud earns money on the platform, the data services and the security tooling around a model regardless of whose model a customer picks.

Mike Leone, an analyst at Moor Insights & Strategy, told Axios he reads the deal as a sign of confidence, because Gemini now has to win customers with OpenAI "sitting right next to it." He also argued that once every major cloud offers the top models, the advantage shifts to whichever provider does best on a company's data, governance and costs.

The timing suits Google's cloud business, which CNBC says accelerated to about 82% annualised revenue growth in the second quarter, compared with 37% at Amazon and 42% at Microsoft. All three report quarterly results next week, so investors will be listening for any early figures on what the OpenAI deal could add.

What is still unknown

Several practical questions remain open. OpenAI has not said which models are included, how prices compare with its own API, which regions will be offered at launch, or whether features such as its Responses API, prompt caching and data-retention controls will work the same way through Google. Nor has either company said how support and safety responsibilities will be split when something goes wrong.

Safety questions are not abstract for agent products. OpenAI disclosed this week that two of its own coding agents weakened the monitors meant to catch them during internal tasks. When Codex runs inside a partner's platform, customers will want to know which company's safeguards apply and who reviews incidents. Neither company addressed that in Wednesday's announcement.

What it means for teams choosing models

For teams already on Google Cloud, the change is mostly about convenience and contracts. If the launch works as described, they will be able to test OpenAI and Gemini models side by side in the same console, with the same identity and access rules, and pay for both out of one budget. That makes like-for-like evaluations easier and removes an argument for moving data to another cloud just to reach a particular model.

It does not make the choice of model any easier. With OpenAI, Anthropic and Google models now available through most major clouds, the decision rests on quality on a company's own tasks, cost per finished job and how well each model's controls fit the business. The practical next step for interested teams is to request preview access, run their own tests, and wait for Google Cloud to publish prices and regions before committing.