OpenAI announced on 21 October 2026 that its frontier models and Codex agentic workflows are coming to Google Cloud Model Garden. If your company already has a GCP spend commitment, prepare an evaluation now but keep production OpenAI traffic on its current route. General availability is planned for the coming weeks.
How can you use the new procurement route?
Google Cloud customers will be able to access OpenAI models through Model Garden, with purchases transacted through Google Cloud Marketplace. Enterprises will be able to draw down their GCP spend commitments for those purchases.
This can matter to an organisation that already runs systems on Google Cloud and wants to buy OpenAI models within the same cloud contract. Subject to the agreement, the Marketplace route may reduce separate vendor approvals and procurement steps. The listing and the existing GCP contract together determine whether that advantage is material.
Customers can already request access. That is enough to prepare the evaluation, test cases and procurement comparison. A production migration should follow the final commercial and technical terms.
OpenAI says Microsoft remains its primary cloud partner.
When does a GCP spend commitment create an advantage?
A GCP spend commitment creates a procurement advantage when you already have such an agreement and purchases of OpenAI models count towards it. Check how your own contract treats Marketplace purchases.
Include the owner of the Google Cloud contract, the technical owner of the AI workflow and the person responsible for data processing in the review. A financial comparison is useful only when Google Cloud Marketplace, the direct OpenAI API and Azure OpenAI are tested on the same task and functional scope.
If you have no GCP spend commitment, the announcement alone is insufficient grounds for signing a new Google Cloud contract. Google Cloud's chief executive says broader model choice can help lower costs. Compare that vendor claim with Marketplace prices and task-level results from your own workload.
The related material on the announcement page includes an earlier page about buying OpenAI models and Codex through an Oracle cloud commitment. This suggests a broader pattern in which OpenAI sells through existing commitments with several hyperscalers. The terms of the Oracle offer should not be applied automatically to the Google Cloud route.
The new channel weakens the assumption that OpenAI models require either a direct contract or Azure. Vendor dependence remains. The model provider, marketplace and cloud infrastructure may each add a contractual layer, so future switching costs still depend on the integration and commercial terms.
Which commercial and technical terms should you compare?
The general availability offer should identify the model range, prices, limits, Codex functions and support arrangements. “Frontier models” is not a substitute for model identifiers, supported tool calls and usage limits. Test feature parity between the direct OpenAI API and the Marketplace offer rather than assuming it.
The announcement in OpenAI's developer community also says the models will soon become available in Google Antigravity and the Gemini Agent. Treat these as separate offers and compare their access methods, functions and contract terms with Model Garden.
A full cost comparison requires the Marketplace price list. Beyond token charges, account for how much a purchase reduces the remaining GCP commitment, which support services are included, and how much integration or migration work the route requires.
| Decision criterion | Statement on 21 October 2026 | What to obtain before purchase |
|---|---|---|
| Availability | General availability is planned in the coming weeks | Launch date, countries and regions |
| Procurement | Purchases will be transacted through Google Cloud Marketplace | Whether your GCP spend commitment is eligible |
| Model range | Frontier models and Codex agentic workflows | Model identifiers, limits and Codex functions |
| Cost | Billing through Marketplace | List price and cost per successful task |
| Contract and support | The transaction runs through Google Cloud Marketplace | Ownership of the SLA, support and incident response |
*Source: OpenAI's announcement of 21 October 2026.*
Run the same real business tasks through all three routes. For teams serving Hungarian users, include Hungarian prompts and documents. If the system works in several languages, test every production language separately. Compare task success, latency, cost per successful task and error handling. Test tool calls with the same functions and permissions on each route.
If you want to keep several model procurement routes open, review the architecture choices behind enterprise AI platforms. Enterprise AI platforms
What must you establish about data processing?
OpenAI says the integration will use Google Cloud's access controls, centralised data-governance controls and infrastructure.
Before production use, obtain written answers covering:
- the EU regions available for each model;
- where inputs, outputs and logs are processed;
- which agreement and data-processing terms govern the Marketplace purchase;
- whether Zero Data Retention or an equivalent arrangement applies;
- which party handles access, deletion and incident requests;
- how the processing chain differs from the direct API or Azure OpenAI.
If the workflow handles personal or commercially sensitive data, these answers should also determine the scope of the pilot. Availability through Model Garden does not by itself establish that data will be stored and processed within the European Union.
What should you check before making a purchase decision?
- Name the workflow on which you will compare the three procurement routes.
- Confirm in writing whether purchases count towards your GCP spend commitment.
- Request the launch date, regions, model identifiers and included Codex functions.
- Obtain comparable offers for Google Cloud Marketplace, the direct OpenAI API and Azure OpenAI.
- Record processing locations, retention terms and Zero Data Retention availability.
- Run the same test cases through all three routes, in Hungarian and in every other production language.
- Assign contractual ownership for the SLA, support and incident response.
- Keep prompts, model identifiers and the integration layer portable so a later channel change remains manageable.
Which organisations should prepare a pilot?
A pilot is worth preparing if your organisation already uses Google Cloud, has a GCP spend commitment and plans to add an OpenAI model or Codex to a defined workflow. Preparation should cover test cases, data requirements, acceptance criteria and the procurement routes being compared.
Keep production on its current route if you have no GCP spend commitment, require confirmed storage and processing within the EU, or cannot operate without defined ownership of the SLA and support. Waiting is also reasonable when the direct OpenAI API or Azure OpenAI already works well and the alternative channel has no quantified procurement advantage.
For the wider Google agent decision, use the comparison of Gemini Agent and custom agent orchestration. Before negotiating a longer cloud agreement, the guide to Maia 300 and Azure AI contract terms provides additional contract questions.