OpenAI models are now on Google Cloud — billed against GCP commitments
For eight years OpenAI and Google were the argument: one company built the transformer, the other built the product on top of it and then tried to eat its lunch. As of today one of them sells the other's models by the token, on the same invoice.
OpenAI's frontier models and its Codex coding agents are coming to Google Cloud's Model Garden, bought through the Google Cloud Marketplace, with general availability promised "in the coming weeks," according to OpenAI's announcement. The lineup Google Cloud itself named is the current top of OpenAI's stack: GPT-6 Astra, GPT-6.1 Sol and GPT-6 Luna, plus Codex agentic workflows, with access also planned inside Antigravity and the Gemini Agent. Microsoft, OpenAI was careful to add, remains its primary cloud partner.
The interesting part is not the logo going up on a partner wall. It is the plumbing underneath it. Google Cloud says purchases are transacted directly through its Marketplace, so OpenAI usage lands on a single Google Cloud invoice, draws down existing cloud commitments and counts toward Flexible Savings Plans across model providers, per Google Cloud's launch post from Michael Gerstenhaber, its vice president of product for Gemini. That is a procurement decision dressed as a partnership announcement, and it is the reason this deal matters more than the four cloud logos it now completes.
The timeline has been compressed to the point of comedy. Less than six months ago OpenAI said Microsoft would no longer hold an exclusive license on its intellectual property. Amazon announced it would host OpenAI models the next day, in April. Oracle followed in June, selling OpenAI models and Codex against Oracle cloud commitments. Google is the fourth channel. OpenAI models now run on all three of the big US clouds, which is a strange place to arrive for a company founded in 2015 partly as a counterweight to Google's concentration of AI research.
The invoice is the moat now
Committed cloud spend is the strongest switching cost in enterprise software, and it is the one thing a better model cannot argue with. A team holding a large, partly unspent Google Cloud commitment now faces a price list where the marginal cost of routing a workload to GPT-6 Astra is effectively zero until the commitment is consumed. The same request, priced per token, costs the buyer nothing extra on the invoice and costs the buyer everything if they pay cash for it out of a separate budget line. That is how model choice actually gets made inside large companies: not by whoever wins the benchmark, by whoever is already paid for.
Both sides get something real out of that. Google Cloud was already the fastest-growing of the three hyperscalers, with annualized revenue growth in the segment accelerating to about 82% in the second quarter from 63% in the first, against 37% at AWS and 42% at Azure, CNBC reported. Now it also collects a marketplace cut on the tokens its own customers route to a rival's model, and it makes the commitment stickier by giving that unspent balance one more thing to buy. For OpenAI, this is distribution into accounts it could not reach on merit alone — the enterprises where procurement, not preference, decides which model a team is allowed to use. That matters while OpenAI is courting investors at a $1.4 trillion valuation and pointing at an initial public offering.
The model names tell their own story too. Astra, Sol, Luna: demanding work, near-Astra at lower cost, fast and cheap for everyday volume. Three tiers on one catalog page, priced for comparison shopping. That is what a product line looks like when the buyer is a platform team filling a cart, not a research team reading a paper. We watched the same shift when OpenAI opened its Codex agent harness to every developer with the Agents API — the model stopped being the product and became the meter.
What the announcement overstates
Marketplace access is not migration. Most large teams already run OpenAI on Azure with provisioned throughput, private networking and a security review that took a year; nobody re-platforms a production workload because a second checkbox appeared on a page. Data gravity does more to keep workloads where they are than any rate card does to move them, and Azure remains the default path for the buyers who want OpenAI and nothing else to think about.
Then there is the part of the announcement that is written in future tense. Google's post lists Zero Data Retention as available "coming soon" and Confidential Compute as coming soon as well. Those two features are the actual gate for regulated buyers — a bank or a hospital cannot put customer data through a model without them, which means what launched today is a pilot path, not a production one. Read the availability date at general availability, not the press release.
And the channel conflict is real, if politely managed. Google is now reselling GPT-6 Astra on the same shelf where it markets Gemini 4 Argon as the better value, and its Marketplace incentive is gross merchandise volume, not loyalty to its own model. Google also claims to be the only hyperscaler offering all three leading closed-source model families. That is a breadth claim, and breadth became table stakes the week AWS started hosting OpenAI in April. Neither announcement published per-token rates; pricing is undisclosed until the rate card ships.
What to watch
All three hyperscalers report earnings next week, and Google Cloud's growth line is the one to read first — a marketplace number that starts moving is the signal that this deal is doing real work rather than decorating a slide. Then the GA date for Model Garden access, whether Antigravity ships Codex with cloud-hosted tasks, and whether AWS answers by matching the commitment mechanics rather than the model list.
If you are the buyer: ask for the rate card before the pilot, check whether the Flexible Savings Plan pass-through applies to your commitment type, and price the same workload against each cloud balance you are sitting on. The same prompt now has a different effective price depending on where your prepaid dollars are stranded, and that spread is worth more than most negotiation cycles.
If your company's AI budget is a cloud commitment, does model choice even matter anymore? Write to us via our contact page.
