What actually launched
Google Cloud announced on October 21 that three OpenAI models, GPT-6 Astra, GPT-6.1 Sol, and GPT-6 Luna, are rolling out in private preview to select enterprise customers through its Vertex AI Model Garden. Access currently requires submitting an interest form rather than self-service signup through the console, and Google's own partner-model documentation had not yet listed the models at the time of the announcement, a gap that suggests the rollout is still being finalized internally even as the press release went out. OpenAI's own messaging said general availability on Google Cloud is planned for the coming weeks, language that leaves real room for slippage given how recently this preview opened.
The three models target different price-performance points for enterprise buyers. Astra is positioned for demanding work such as coding, computer use, and scientific research where raw capability matters most. Sol delivers close to Astra-level capability at a meaningfully lower cost per token for teams that do not need the top tier. Luna is built for fast, high-volume, cost-efficient tasks at scale, the kind of workload where margin per call matters more than peak intelligence. That segmentation mirrors how OpenAI already sells these same models elsewhere, so the actual news here is distribution reach, not a new model lineup or new capability.
The six-week gap that matters more than the launch
Microsoft Foundry made GPT-6 Astra generally available on September 3. AWS Bedrock followed on September 8, with both platforms also listing Sol and Luna as fully generally available, not merely in preview. Google Cloud is landing roughly six to seven weeks later, and only in preview status. For a model family this prominent, that gap is a genuine signal about how OpenAI prioritized its distribution partners this cycle, and quite possibly about commercial terms Google and OpenAI needed longer to work out than either Microsoft or Amazon required.
Google's consolation is a claim no competitor can currently match: it is the only hyperscaler that natively offers Gemini, Anthropic's Claude, and now OpenAI's GPT-6 on a single platform. That claim holds up in a narrow access sense, but it remains a weaker selling point while GPT-6 sits in preview with no published pricing, no confirmed regions, and features like zero data retention and confidential compute both listed as coming soon rather than actually available to customers today.
How the billing actually works
Purchases run through Google Cloud Marketplace and land on a single Google Cloud invoice, which is the operational convenience most enterprises actually want from this kind of integration: one vendor relationship, one procurement cycle, one invoice to reconcile each month instead of juggling a separate OpenAI billing relationship on top of existing cloud spend. Usage draws down existing Google Cloud spend commitments, which sounds like a discount mechanism but functions quite differently in practice. Google's own Flexible Savings Plan documentation excludes third-party products from FSP discounting entirely, so the real benefit is commitment drawdown against existing spend, not a lower effective per-token rate on GPT-6 usage.
That distinction will matter the first time a finance team reconciles actual GPT-6 costs against an annual budget built on assumed discount rates. Teams that assumed their committed-use discount applies across any model running on the platform will find it does not extend to OpenAI's models at all, and should model GPT-6 usage at its full undiscounted rate until Google explicitly states otherwise in updated pricing documentation, which procurement should request in writing before any production rollout.
What is still missing
Google says prompts, context, and outputs stay within the customer's data perimeter and are not used to train Google's own foundation models, a baseline commitment most enterprises will reasonably expect by default at this point in the market. Zero data retention and Confidential Compute, both already available in some form through competing platforms, remain listed as coming soon rather than shipped on Google's GPT-6 integration. No regions have been named either, which makes any firm data residency or latency commitment effectively impossible to pin down today for a procurement team trying to plan ahead.
One genuinely useful detail worth noting: Google says customers can connect local Codex CLI and Desktop workflows to its Gemini platform for inference starting the day of the announcement. That capability is not exclusive to Google, however. AWS announced similar Codex support for Bedrock back in September, so developers already working across multiple clouds will find the workflow familiar and incremental rather than genuinely novel when they try it on Google's side.
The call for enterprise buyers
If your organization is already standardized on Google Cloud and has specifically been waiting for GPT-6 access to avoid a separate multi-cloud integration project, this preview is worth testing now, precisely because it lets you validate model behavior and latency characteristics before committing any production traffic to it. Use the preview period to build your own internal cost model against the undiscounted rate, since that is almost certainly what you will end up paying at general availability unless Google changes its Flexible Savings Plan policy in the meantime.
Hold anything regulated or customer-data-sensitive until Google publishes zero data retention terms, confirmed regions, and a rate card you can actually underwrite with confidence. The six-week gap behind Microsoft and AWS is a reasonable proxy for how much slower Google's commercial terms for this particular integration are likely to mature, and rushing a production commitment ahead of that maturity curve is the kind of unforced error that tends to surface in next year's cloud cost audit.
