GPT, Claude or Gemini: Which Cloud Sells Which Model

Google Cloud announced on October 21, 2026 that OpenAI's GPT-6 Astra, GPT-6.1 Sol and GPT-6 Luna were rolling out in private preview in its Model Garden, and said in the same post that it was now the only hyperscale cloud provider offering access to all three leading closed-source model families. That is a checkable claim, and so is its mirror image: which of the three big clouds sells which frontier model today, under what label, since when, and on whose invoice. This page is the check. Every cell below comes from Amazon Bedrock's model cards, Microsoft Foundry's catalogue pages or Google Cloud's Model Garden docs, read on October 21, 2026, with the launch blog posts used only for dates and quotes. What Google Cloud customers gain from the OpenAI deal, and when, is in our OpenAI-on-Google-Cloud post; this one is the matrix.

The short version. Google's sentence holds on its own terms: it is the only one of the three whose catalogue lists Gemini, and it also lists Claude and, in private preview, GPT-6. It is also the only one of the three where GPT-6 is not yet generally available. AWS and Azure each sell four GPT-6-series models as GA and the Claude 5.5 family as GA, and neither lists a single Gemini model. A cloud buyer with commitments on one provider has a very different menu from a buyer on another, and the labels, GA versus preview versus gated, matter as much as the names.

The sentence under test was written by Michael Gerstenhaber, Google Cloud's VP of Product Management for Gemini, in the October 21 announcement. It is a specific claim about three vendors and three clouds, and it can be read against the catalogues.

On the pages we read, it holds. Bedrock's models-at-a-glance page lists the GPT-6 family and the Claude 5.5 family, and its Google row contains Gemma models only; the word Gemini does not appear. Foundry's two catalogue pages list GPT-6 and Claude and no Gemini model either. Google's own docs list Gemini 4 Argon and Gemini 3.8 Flash as GA and Claude 5.5 as GA, and its blog puts GPT-6 in private preview. So "all three" is true of Google and of nobody else, with two qualifiers the sentence does not carry: Google's GPT-6 access is a private preview that the public docs did not list on October 21, while the other two clouds have sold it as GA for six to seven weeks; and "access to" is doing work, because a private preview for select enterprise customers is not a catalogue entry anyone can enable. The example of the gap is the Codex routing Google announced the same day: AWS's September 8 post already says customers can configure ChatGPT Work and Codex to use GPT-6 Astra on Bedrock, and Microsoft's Foundry docs, dated September 3, set up the Codex CLI with gpt-6-astra on Azure OpenAI, so Google's version is a feature the other two big clouds already had, not a first.

Labels are the clouds' own. GA means the page says generally available, or, on AWS, shows a model lifecycle of Active with no preview notice. Preview, private preview and gated are the page's words. Not listed means the catalogue pages we read do not show the model, which is different from not available: none of the three clouds says that about any model here, and Azure's docs say they show a selection. Model identifiers are given where they matter for telling two entries apart.

The dates settle the order. Microsoft's blog of September 3, 2026 announced GPT-6 Astra as generally available for all customers in Microsoft Foundry, and its September 22 post added Sol and Luna to the GA lineup, followed by GPT-6.1 Sol on September 29. AWS's What's New post of September 8 announced general availability of GPT-6 Astra on Amazon Bedrock, and its model cards list Sol and Luna with a September 22 launch and GPT-6.1 Sol with September 29, all with an Active lifecycle and AWS Marketplace product identifiers. Google's October 21 post names Astra, 6.1 Sol and Luna in private preview for select enterprise customers, with more coming soon, and OpenAI's companion post says general availability is planned in the coming weeks. The Model Garden partner-models page, last updated October 21, did not list any OpenAI model, and the docs path where an OpenAI page would sit returned a 404.

Two details that only the catalogues show. Both AWS and Azure still list the original GPT-6 Sol alongside GPT-6.1 Sol, so each sells four GPT-6-series models where Google names three. And the open-weight gpt-oss models are GA on Bedrock without a Marketplace subscription, in preview on Foundry, and listed with no launch stage on Google. The pricing that goes with these rows, where published, is in our frontier model API price index; Google has published no GPT-6 price for its route, and this page prints none.

Claude is the family with the cleanest row. Opus 5.5, Sonnet 5.5 and Haiku 5.5 are GA on Bedrock with regional profiles for the US, EU, Australia and, for Opus and Haiku, Japan, GA on Foundry, and GA on Model Garden with release dates of September 22, September 28 and October 7, 2026 that match Bedrock's launch dates to the day. Azure adds a choice the others do not: each Claude model is offered hosted on Azure or hosted on Anthropic's infrastructure outside Azure, both GA, with Anthropic as the seller and operator and an independent data processor under either option, and a fast variant of Opus 5.5 hosted on Azure.

Mythos is the opposite. Bedrock lists Claude Mythos 5.1 as a preview gated to a vetted set of organisations through Anthropic's trusted access programmes and offered as a beta service; Foundry lists Mythos 5.1, 5 and a preview build as a gated research preview on Anthropic infrastructure only; Google's Claude catalogue does not list it at all, although its abuse-monitoring and retention terms name Mythos, so some arrangement exists that the catalogue does not describe. For how the three clouds differ when Claude Code is pointed at them, see our Claude Code on Bedrock, Vertex and Foundry guide.

Gemini 4 Argon, GA on Model Garden since October 13, 2026, and Gemini 3.8 Flash, GA since September 2, appear on neither of the other clouds. Bedrock's Google row holds Gemma 4 and Gemma 3 only; Foundry's pages hold no Google model at all. That is the structural fact behind Google's claim: Microsoft and Amazon can resell OpenAI's and Anthropic's models because those labs sell through every cloud, and Google does not.

The rest of the matrix is where buyers get surprised. xAI's Grok 4.7 is GA on Bedrock and on Foundry, where Azure files it under "SpaceXAI models sold by Azure", but only a preview on Google, added September 30. Meta's Muse Spark 1.3 is a preview on Google alone, reached through an account manager, and is listed on neither Bedrock nor Foundry, which carry Llama only; Muse Spark Open is on none of the three. Mistral Large 3 is GA on Bedrock, a preview on Foundry, and not listed on Google, which carries Medium 3, Small 3.1, Codestral 2 and OCR instead. DeepSeek is V3.2 on Bedrock, V4 on Foundry and V4.1 Flash on Google. Qwen has a full inference lineup on Bedrock and Google but appears on Foundry only in a fine-tuning table. Kimi K3 is GA on Bedrock, K2.7 and K2.6 are previews on Foundry, and Google lists K2 Thinking. Our Grok on Bedrock guide covers how a non-OpenAI, non-Anthropic model lands on a hyperscaler.

For a buyer with a cloud commitment, who sells the model decides whether the spend counts. On Bedrock, OpenAI and Anthropic models are sold by their providers as third-party content under AWS's service terms and show up as an AWS Marketplace bill, with the subscription started automatically on first invocation; Amazon, DeepSeek, Google, Meta, Mistral, Moonshot, Qwen, xAI and others are not sold through Marketplace. The GPT-6 Astra card states that in-region and geographic cross-region prices carry a 10% premium over OpenAI's own rate for the same tier, $11 input and $55 output per million tokens for short context, while global routing matches OpenAI's price. On Azure, GPT-6 is sold by Microsoft and billed through the Azure subscription under Azure's service agreements, at $10 and $50 for global standard; Claude is a partner model that needs an Azure Marketplace subscription, bills in Claude Consumption Units, and the CCU meter is eligible for a Microsoft Azure Consumption Commitment. On Google, OpenAI usage is transacted through Google Cloud Marketplace onto a single invoice, draws down existing commitments and counts toward Flexible Savings Plans; Google's own FSP page adds that third-party products get no discount from the plan and that drawdown follows the Marketplace rules and caps in the enterprise agreement.

All three say a version of the same first sentence, with Claude exceptions noted below: your content is not used to train the base models and is not shared with the provider. The differences are in retention and in what is available today. AWS says content is not shared with any model provider, that providers have no access to deployment accounts or logs, that classifier-flagged traffic to OpenAI models is retained up to 30 days for abuse detection with full zero data retention available to eligible customers on request, and that an account or project can set a zero-retention mode that blocks any model requiring retention. Microsoft says prompts and completions to Azure-sold models are not available to OpenAI and not used to train without permission, with modified abuse monitoring by application. Google says data stays in the customer's perimeter and is never used to train, and that ZDR options and Confidential Compute for OpenAI models are coming soon, which means neither was available on October 21. For Claude Fable 5 and Mythos 5 on Google, prompts are retained up to 30 days and customers must enable sharing with Anthropic for abuse monitoring.

A matrix that hides its gaps teaches readers that the gaps are facts. These are not. Google's public docs did not list any OpenAI model on October 21, so model identifiers, regions, quotas and prices for that route could not be confirmed; the preview catalogue needs a signed-in console. Only OpenAI's post gives a timing for Google GA, and it says weeks. No AWS page we read says whether Bedrock Marketplace spend draws down an enterprise commitment, and the only explicit MACC statement on Azure covers Claude, not Azure OpenAI. Google's FSP page covers participating third-party models without listing them, so whether Claude on Model Garden counts is unconfirmed; the explicit drawdown sentence covers OpenAI. Whether Mythos can be reached on Google at all is unknown. Per-model region availability was not compiled, and the Foundry portal and Bedrock Marketplace self-hosted endpoints were not checked, so "not listed" applies to the doc pages only.

The question this table answers for a buyer is which cloud's commitment can pay for which model this quarter. How those commitments interact with OpenAI's and Anthropic's own enterprise terms is the subject of our marketplace spend commitments comparison, and choosing a route for a production agent platform is part of what our AI transformation engagements cover.