--- title: "OpenAI on Google Cloud: Three GPT-6 Models, No Price List Yet" description: "GPT-6 Astra, GPT-6.1 Sol and GPT-6 Luna are coming to Google Cloud's Model Garden. What is confirmed, what Google has not published yet, and the router rule OpenAI added the same day." author: "Jahanzaib Ahmed" date: 2026-10-22 category: "implementation" readingTime: "8 min read" tags: ["ai news", "openai", "google cloud", "llm pricing", "ai-agents"] canonical: https://www.jahanzaib.ai/blog/openai-google-cloud-model-garden source: https://www.jahanzaib.ai --- # OpenAI on Google Cloud: Three GPT-6 Models, No Price List Yet OpenAI Google Cloud access is real, but only for a short list of enterprises. On October 21, 2026, OpenAI announced that [GPT-6 Astra, GPT-6.1 Sol and GPT-6 Luna are coming to Google Cloud's Model Garden](https://openai.com/index/openai-models-google-cloud/), billed through Google Cloud Marketplace and payable from existing Google Cloud spend commitments. OpenAI says general availability is planned "in the coming weeks," and Google's own post, dated October 22, says the models are rolling out in private preview now. Google now joins Microsoft, AWS and Oracle as a cloud selling OpenAI's frontier models. And it arrived on the same day OpenAI published a usage policy change that touches anyone running a [model router or AI gateway](https://www.jahanzaib.ai/blog/llm-router-ai-agent-routing) in front of its models. For builders, that fine print beats the headline. ## What the OpenAI Google Cloud deal includes [Google's post](https://cloud.google.com/blog/products/ai-machine-learning/google-cloud-brings-openai-models-to-customers), written by Michael Gerstenhaber, VP of product management for Gemini, lists three models: Astra for demanding coding, computer use and research work, Sol for near Astra quality at lower cost, and Luna for fast, cheap everyday volume. Usage lands on a single Google Cloud invoice, draws down existing commitments and counts toward Flexible Savings Plans. 3OpenAI models in private preview on Google Cloud ([Google Cloud blog](https://cloud.google.com/blog/products/ai-machine-learning/google-cloud-brings-openai-models-to-customers)) 200+models already in Model Garden, per [Google](https://cloud.google.com/blog/products/ai-machine-learning/google-cloud-brings-openai-models-to-customers) $10 / $50GPT-6 Astra per million input and output tokens on OpenAI's own API ([pricing page](https://developers.openai.com/api/docs/pricing)) Nov 20when OpenAI's updated usage policies take effect ([OpenAI](https://openai.com/index/broad-access-through-consistent-standards/)) Developers in the preview can call the models through Google's APIs today, and Google says local Codex CLI and Codex Desktop setups can point at its platform for inference. Antigravity, Google's agent based coding tool, and Gemini for business get the models "coming soon." OpenAI's post ends with: "Microsoft remains OpenAI's primary cloud partner." On data handling, Google says prompts and outputs stay inside the customer's data perimeter and are never used to train foundation models. The models run on dedicated Google Cloud infrastructure, behind VPC Service Controls and unified audit logging. Two items are still marked coming soon: Zero Data Retention options and Confidential Compute. ## Google's docs still lack model IDs, regions and prices I went looking for the details I would need to wire this into a production agent, and they are not published. The [OpenAI models page in Google's documentation](https://docs.cloud.google.com/gemini-enterprise-agent-platform/models/maas/openai), last updated October 21, still lists only the open weight gpt-oss 120B and gpt-oss 20B. There is no GPT-6 model ID, no region table, and no price. Note the URL, too: Vertex AI documentation now lives under the name Gemini Enterprise Agent Platform. So the facts that decide whether you can move a workload are missing: regions, Google's price, and which Responses API features come along. Until Google publishes a model card, anyone outside the preview is planning on press release language. **Note:** "Private preview to select enterprise customers" means access goes through Google's interest form. ## Prices to budget against until Google publishes its own OpenAI's direct API prices are the only firm numbers. These are per million tokens, standard tier: | Model | Input | Cached input | Output | | --- | --- | --- | --- | | GPT-6 Astra | $10.00 | $1.00 | $50.00 | | GPT-6.1 Sol | $2.00 | $0.10 | $10.00 | | GPT-6 Luna | $0.10 | $0.01 | $0.50 | Source: [OpenAI API pricing](https://developers.openai.com/api/docs/pricing). That page also says OpenAI's regional processing (data residency) endpoints carry a 10% uplift for models released on or after March 5, 2026. Sol costs a fifth of Astra and Luna a hundredth, which matches [SiliconANGLE's read of the lineup](https://siliconangle.com/2026/10/21/openai-to-make-its-most-advanced-ai-models-available-in-google-cloud/). My read: budget at OpenAI's list price and treat any Google discount as a bonus. For most companies the saving is not per token at all. It is the commitment. If your finance team signed a large Google Cloud commit and you are behind on burning it, routing Sol or Luna traffic through Model Garden spends money you already owe. If you have no commit, there is no obvious reason yet to switch billing away from OpenAI. I made the same argument about cached input when [GPT-6.1 Sol halved its cache rate](https://www.jahanzaib.ai/blog/gpt-6-1-sol-pricing-prompt-caching-migration): on an agent loop, the cached input line decides the bill. Whatever Google publishes, check that row first, and check that [prompt caching](https://www.jahanzaib.ai/glossary/prompt-caching) works the same way through its endpoint. ## Bedrock already shows what a cloud copy leaves out The best evidence of what to expect sits in OpenAI's own docs. [GPT-6 Astra went generally available on Amazon Bedrock on September 8](https://aws.amazon.com/about-aws/whats-new/2026/09/openai-gpt-6-astra-on-amazon-bedrock/), and OpenAI maintains a [feature matrix comparing its API with Bedrock](https://developers.openai.com/api/docs/guides/amazon-bedrock). The same model ID does not buy the same API. | Responses API feature | OpenAI API | Amazon Bedrock | Google Cloud | | --- | --- | --- | --- | | Function calling, structured outputs, streaming | Available | Available | Not documented | | Prompt caching | Available | Available | Not documented | | Remote MCP servers | Available | Not available | Not documented | | Hosted file search | Available | Not available | Not documented | | Shell tool | Available | Not available | Not documented | | Multi agent (beta) | On supported models | Not available | Not documented | | WebSockets and mid turn steering | Available | Not available | Not documented | | Hosted web search | Available | Mantle endpoint only | Not documented | ![Diagram of an agent passing a feature check before choosing the OpenAI API, Google Cloud or Amazon Bedrock](https://cdn.sanity.io/images/qajb7q5q/production/cf2c4c0a557747406787a4645049f75603db9ba2-1376x768.png?w=1200&q=75&auto=format&fit=max) _The model name is the same on every path; the tools your agent can call are not, so the feature check comes before the billing decision._ If your agent talks to an [MCP server](https://www.jahanzaib.ai/glossary/mcp-server) through OpenAI's hosted remote MCP tool, or leans on hosted file search for retrieval, a Bedrock port means rebuilding those pieces in your own code. Expect Google's version to have its own gaps. When I evaluate a move like this, I list every tool type in the request payload first, then check each one against the target cloud's matrix. Price comes after. ## OpenAI's November 20 rule for AI gateways On the same day, OpenAI posted [Broad access through consistent standards](https://openai.com/index/broad-access-through-consistent-standards/). It says partners that host or resell its models must keep safeguards consistent with OpenAI's own, investigate violations and coordinate on enforcement. The line that matters: the updated Usage Policies, effective November 20, 2026, require "services that provide direct access to our models such as model routers or AI gateways to get our review and approval before offering that access." That reads as aimed at services that resell access to others, the category I covered when [Stripe bought OpenRouter](https://www.jahanzaib.ai/blog/stripe-openrouter-acquisition-llm-routing). A gateway you run inside your own app to switch between providers is a different thing, but the post does not draw that line. If you run a hosted gateway that other teams or customers call, read the new policy text the day it publishes and decide whether you need to apply. The post also introduces SafetyIDs, a way for customers to tag end users so OpenAI can investigate misuse by one account without learning who that person is. For any multi tenant product, that is worth adopting early. One leaked key or abusive user should not get your whole organization restricted. ## When to buy OpenAI through Google Oracle went first: OpenAI announced Oracle access on June 10 with the same "coming weeks" wording, and the [Oracle Marketplace listing went live on August 7](https://openai.com/index/openai-on-oracle-cloud/), United States only. Roughly eight weeks from announcement to purchase. AWS followed with GPT-6 Astra on Bedrock in September. My earlier take on [what OpenAI on AWS meant for agent builders](https://www.jahanzaib.ai/blog/openai-on-aws-bedrock-managed-agents-2026) holds up: the cloud listing is about who signs the invoice. Google Cloud is the better home for OpenAI models if: - You have an unspent Google Cloud commitment and steady Sol or Luna volume. - Your data already sits in BigQuery or Cloud Storage and your security team will not approve a new vendor path. - You want Gemini, Claude and OpenAI models behind one IAM, one audit log and one bill. Google says it is now the only hyperscaler with all three closed model families. - Your agent uses plain function calling and structured outputs, not hosted OpenAI tools. Stay on OpenAI's API if your agent depends on remote MCP, hosted file search, the shell tool or WebSocket streaming, or if you need a new model the day it ships. Cloud copies trail: OpenAI's Bedrock guide still lists GPT-6 Sol, not the newer GPT-6.1 Sol. For the bigger question of which model family fits a business build, my [OpenAI versus Claude comparison](https://www.jahanzaib.ai/blog/openai-vs-claude-business-ai-agents) is the longer answer. ## Steps to take before general availability 1. **Inventory your tool use.** List every tool type your OpenAI requests send, hosted and custom. Every hosted tool on it is work you redo. 2. **Ask finance one question.** How much Google Cloud commitment is unspent this year? If the answer is zero, this news changes little for you. 3. **Request preview access only with a test plan.** Run the same eval set through OpenAI direct and Model Garden, and compare outputs, latency and the cached input bill. 4. **Watch three pages:** Google's OpenAI model doc for IDs, regions and prices; the Zero Data Retention status; and OpenAI's Usage Policies on November 20. 5. **Keep the model ID configurable.** OpenAI is [retiring older models on October 23](https://www.jahanzaib.ai/blog/openai-model-shutdown-october-23), and cloud copies follow their own schedules. Hardcoded IDs break twice. If you would rather have an agent built so the provider is a config change rather than a rewrite, that is the kind of system I build: [see how I build AI agents](https://www.jahanzaib.ai/agents). ## Frequently asked questions ### Can I use OpenAI models on Google Cloud today? Only if you are in the private preview. Google says GPT-6 Astra, GPT-6.1 Sol and GPT-6 Luna are rolling out to select enterprise customers through an interest form. OpenAI says general availability is planned in the coming weeks. Oracle's similar launch took about eight weeks from announcement to a live marketplace listing, so plan for weeks, not days. ### How much do OpenAI models cost on Google Cloud? Google has not published prices. Its documentation page for OpenAI models, last updated October 21, lists only gpt-oss. OpenAI's own API charges $10 and $50 per million input and output tokens for Astra, $2 and $10 for Sol, and $0.10 and $0.50 for Luna. Budget at those rates until Google posts its own. ### Can I pay for OpenAI models with my Google Cloud commitment? Yes. OpenAI says purchases go through Google Cloud Marketplace and can draw down existing Google Cloud commitments. Google adds that usage appears on a single Google Cloud invoice and counts toward Flexible Savings Plans across model providers. ### Does the Google Cloud version support all OpenAI API features? Google has not said. The closest guide is OpenAI's Amazon Bedrock matrix, where remote MCP servers, hosted file search, the shell tool, multi agent and WebSockets are not available. Check every tool your agent calls against Google's model card once it is published before moving production traffic. ### What changes for AI gateways on November 20, 2026? OpenAI's updated Usage Policies take effect that day. They require services that provide direct access to OpenAI models, such as model routers or AI gateways, to get OpenAI's review and approval before offering that access. The announcement does not say whether a gateway used only inside your own product counts, so read the published policy text. ## Related - [OpenAI Retires o1, GPT-4 Turbo and Other Models on October 23](https://www.jahanzaib.ai/blog/openai-model-shutdown-october-23) - [GPT-6.1 Sol Halves the Cache Rate and Leaves Every Other Price Alone](https://www.jahanzaib.ai/blog/gpt-6-1-sol-pricing-prompt-caching-migration) - [The Medicare Portal Told OpenAI's Agent No. The Agent Treated It as a Retry.](https://www.jahanzaib.ai/blog/openai-agent-medicare-breach-ai-agent-incident-response) --- Canonical HTML version: https://www.jahanzaib.ai/blog/openai-google-cloud-model-garden