AivexaNewsSearch
AI news for builders and product teamsChecked every hour

OpenAI's latest AI models are coming to Google Cloud and Antigravity

Collected Oct 21, 2026

OpenAI and Google have reached an arrangement that will bring OpenAI's frontier models onto Google Cloud and into Antigravity, Google's agent-powered coding tool. Google says all three current OpenAI models are rolling out now in private preview for select enterprise customers, with broader availability in the Model Garden expected in the coming weeks. OpenAI simultaneously noted that Microsoft remains its primary cloud partner, so this is an expansion of where its models run rather than a change of primary allegiance.

The models involved are the current GPT 6 and 6.1 lineup, which the arrangement names as Astra, Sol and Luna. All three are covered, and Astra is described as the most expensive of the trio to run. For enterprise buyers, the models will appear inside Google's Model Garden, the catalog that already lists more than 200 options spanning Google's own Gemini models, open-weight families such as Mistral, Llama and Gemma, and Anthropic's Claude, which has been available there for some time. Google's claim attached to the deal is that it is now the only large cloud provider offering access to all three leading closed LLMs.

Operationally, the OpenAI models will run entirely on Google's infrastructure, with transactions handled through the Google Cloud Marketplace. That marketplace detail matters more than it might appear: enterprises commonly hold committed Google Cloud spend, and billing through the marketplace lets them apply existing commitments to GPT usage rather than opening a separate procurement channel. In practice this lowers the friction of trying GPT models in place of whatever a team currently uses, since the purchase can be absorbed into an existing cloud budget line.

Access for early users works through a request process. Google says interested customers can submit a request for early access, which enables use through the API even if they are working in Codex CLI or Codex Desktop; only the inference itself runs on Google's systems. Developers already using OpenAI's coding tools would therefore be able to keep their client-side workflow and simply change where the model calls are served.

Google also describes a routing capability on its system that can dynamically move workloads between LLMs based on performance, latency and efficiency. This suggests a setup where a single application can direct requests across multiple models rather than committing to one. A likely trade-off is that such routing adds a layer of abstraction and variability: teams gain the ability to shift load away from a slow or costly model, but they also take on the work of defining what performance, latency and efficiency thresholds should trigger a switch, and of verifying that outputs stay consistent when requests land on different models. Model behavior can differ in ways that are not captured by aggregate benchmarks, so routing has to be tested carefully before it is trusted in production.

The Antigravity piece is the part individual developers will notice most. The coding client has so far supported Google's Gemini and Anthropic's Claude; GPT models are now set to appear there as well. The natural expectation is that the same Astra, Sol and Luna trio will be offered, though that has not been specified. If it holds, Antigravity becomes a single client where a developer can pick among the three major closed model families, which is a meaningful shift for a tool that previously only mixed in one outside vendor.

The timing sits alongside Google's recent rollout of Gemini 4 Argon across its AI footprint, which Google presents as bringing its flagship back to parity with competitors. The reporting notes it appears to be an improvement over Gemini 3.1 Pro but consumes tokens at a very high rate. That gives the OpenAI integration a specific purpose: for teams that find Argon's token appetite costly, Claude and now GPT models provide alternatives inside the same cloud and, potentially, the same coding client. This is my own reading of how the pieces fit, not a stated Google position.

Why it matters: Enterprises with committed Google Cloud spend can soon bill OpenAI inference through the marketplace they already use, and developers may get a choice of three closed model families inside one coding tool. The main consequence to watch is portability: when the same client and cloud can serve Gemini, Claude and GPT, switching models becomes a configuration decision rather than a migration, which raises the competitive pressure on price, latency and token efficiency. Whether that plays out depends on how broadly the private preview opens and whether the Antigravity support covers all three models, neither of which is settled by the announcement.

Read at Ars Technica · AI

Based on reporting from the original publisher. Visit the source for full context and later updates.

Publisher excerpt

Google Cloud will have GPT 6.1 Astra, Sol, and Luna in the coming weeks.