Photo: Pexels
Two of the biggest names in artificial intelligence moved into the same building this week, and neither of them owns the building. On October 21, Google Cloud announced that OpenAI's newest models were coming to its Model Garden. A day later, on October 22, Anthropic said its full Claude platform was generally available through Google Cloud. For anyone who buys, builds with, or simply follows AI, that 48-hour sequence matters more than any single model release this month, because it changes where frontier AI is bought, not just which model is newest.
It is easy to read this as another round of press releases. It is not. Until recently, where you bought a frontier model was largely decided for you: OpenAI meant Microsoft's cloud, Claude meant Anthropic's own platform or a cloud partner's managed version, and Google's cloud sold you Google's models first. This week broke that neat sorting. Google Cloud now sells access to all three of the leading closed model families on one platform, in its own words, alongside more than 200 first-party, third-party and open models. The question worth asking is not who has the best model this week. It is what happens when the shop, not the lab, becomes the battleground.
What actually happened in 48 hours
The OpenAI announcement came first. Google Cloud said GPT-6 Astra, GPT-6.1 Sol and GPT-6 Luna would roll out in private preview to selected enterprise customers through Model Garden, its catalogue for enterprise AI models. OpenAI's own post said the same thing in plainer commercial terms: its frontier models and Codex agentic workflows were entering Model Garden, bought through Google Cloud Marketplace. General availability is planned for the coming weeks, according to OpenAI. An interest form exists for everyone else, and submitting it does not guarantee access.
Anthropic's announcement landed the next day and was different in kind. This was not a preview. Anthropic said its managed Claude platform was generally available through Google Cloud from October 22. Customers sign in with Google Cloud credentials, pay on their Google Cloud bill, and can draw down existing Google Cloud spending commitments. The models on offer through this route are Claude Fable 5.1, Opus 5.5, Sonnet 5.5 and Haiku 5.5, according to Anthropic's launch coverage, with new Claude platform features arriving on this route the same day they reach Anthropic's own platform.
Two other announcements from the same week belong in the same picture, even though they drew less attention. On October 22, Anthropic handed its Agent Skills standard to the Agentic AI Foundation, a Linux Foundation-directed body. Anthropic's figures, as summarised in the week's round-ups, put that standard at more than 70 platforms, around one million skills and roughly 280 million installs. The same day, OpenAI's Decisions API reached general availability: a system built on GPT-6 Luna that returns bounded answers, probabilities, choices or scores, for up to 400 questions in a single request, priced at about $0.10 per million input tokens with no charge for output tokens. And on October 20, the US Department of Homeland Security published an exclusion order requiring federal civilian agencies to remove DeepSeek and affiliated products, including third-party products built with DeepSeek. The Defence Department sits outside that order. Taken together, the week was about distribution and rules, not a new flagship model.
The three OpenAI models, in plain terms
Google Cloud describes the three OpenAI models as a ladder, and the ladder is the point. GPT-6 Astra is positioned for demanding work: coding, computer use, professional tasks and scientific research. GPT-6.1 Sol is described as offering near-Astra intelligence at lower cost, aimed at agentic coding, computer use and professional work. GPT-6 Luna is the fast, cheap option for everyday work at scale. Those are Google's descriptions, not independent benchmark results, and they should be read that way.
For a sense of the price gaps, OpenAI's own API list prices, as summarised by independent coverage of the announcement, are $10 per million input tokens and $50 per million output tokens for Astra, $2 and $10 for Sol, and $0.10 and $0.50 for Luna. Google has not published its own preview prices for these models, so nobody should assume the Google Cloud bill will match those numbers until an account team confirms it. The practical lesson is simpler: the spread between the top and bottom of the ladder is roughly a hundredfold on input price. Choosing the wrong rung for a job is now a budget decision, not a technical footnote.
Codex matters here too. From launch, customers in the preview can connect local Codex command-line and desktop workflows to Google's Gemini platform for inference. OpenAI models in Google Antigravity, its coding tool, and in the Gemini Agent are listed as coming soon rather than live. That split, live versus coming soon, is worth keeping in mind, because announcements tend to blur it and procurement teams pay for the blur later.
Claude's side: two doors into the same cloud
Anthropic's move needs more care than OpenAI's, because Google Cloud now offers Claude through two doors that look alike on an invoice and differ underneath. The new, Anthropic-managed route is operated by Anthropic: Anthropic is the data processor, it sets usage limits, and it runs the service. The older, Google-managed route is operated by Google within its Agent Platform, using the AnthropicVertex software client, and Google is the data processor there.
Anthropic's own guidance draws the line plainly. The Google-managed option remains the better choice for customers who need Google Cloud as the data processor, who have strict regional data residency requirements, or who want native Google Cloud integrations such as Model Armor or Agent Engine. The Anthropic-managed option is the route for customers who want the full Claude platform, including features and betas on the day they launch, with Google Cloud sign-in and billing. Claude Managed Agents, for running longer tasks without customers managing the infrastructure, remains in beta on the new route. Opus 5.5 fast mode is generally available there, listed at $8 per million input tokens and $40 per million output tokens.
One administrative detail deserves a warning of its own: Anthropic says existing discounts must be sorted with an Anthropic or Google Cloud account executive before starting on the Anthropic-managed route, because discounts cannot be applied retroactively to usage. That is the kind of sentence that costs real money if it is discovered after a migration, and it is exactly the kind of sentence that gets skipped in launch-week excitement.
Why Google Cloud, and why now
The timing is not an accident. OpenAI's exclusive distribution arrangement with Microsoft ended in April 2026, and OpenAI has been widening its sales channels since. OpenAI itself still describes Microsoft as its primary cloud partner, and that sentence appeared in its Google Cloud announcement too. Primary does not mean only, and that distinction is the whole story.
For Google, the commercial logic is straightforward. Model Garden already listed more than 200 models, but GPT was the one major closed model family its customers could not buy there. Closing that gap lets Google sell itself as the neutral shelf: buy any leading model, pay one invoice, draw down one spending commitment, and count usage toward Flexible Savings Plans across model providers. For an enterprise that has already committed hundreds of millions to Google Cloud, being able to spend that commitment on OpenAI or Claude models instead of signing a separate contract is a genuine change in purchasing power, not a marketing line.
For OpenAI and Anthropic, the logic runs the other way. Both labs are spending enormous sums on compute, and both need enterprise revenue that arrives through channels customers already trust. Selling through a cloud provider's marketplace puts their models in front of procurement teams who already have a Google Cloud relationship, a security review for Google Cloud, and a budget line for Google Cloud. It also puts the models where the customer's data already lives, which removes one of the quietest objections in enterprise AI sales: the new vendor, the new contract, the new security review.
The quieter moves: skills and decisions
The Agent Skills donation may turn out to be the most durable news of the week. Agent Skills are reusable instructions that teach an AI agent how to do a specific job, from following a company's expense policy to running a particular software workflow. By handing the standard to a foundation rather than keeping it as a company asset, Anthropic is betting that an open, shared library of skills helps its own platform more than a walled garden would. If 70 platforms already use the standard, that bet has a head start. The risk, and it is a real one, is that a foundation moves slower than a single company, and standards that move slowly get routed around. The next test is simple to state: do new skills keep arriving at the same pace under foundation governance, or does the library stall?
The Decisions API points in a different direction. Most AI products return text, and text has to be read, judged and acted on. A decisions API returns a structured answer: a probability, a choice, a score. That is a smaller, duller product, and that is its appeal. A bank deciding which transactions to review, a retailer deciding which items to restock, a support team deciding which tickets to escalate: these are jobs where a number with a confidence attached is more useful than a paragraph. Pricing it at about $0.10 per million input tokens with no output charge also signals who OpenAI is courting here: high-volume, low-drama workloads where the cost per decision has to round to nearly nothing.
What this really means: three connected shifts
First, the cloud is becoming the shop, and the shop sets the terms. When models were sold mainly by their makers, the maker set the price, the contract and the roadmap. When the same models are sold through a cloud marketplace, the cloud provider sets the invoice, the commitment arithmetic and the identity system. That shift favours buyers who already have cloud leverage, because their existing commitment becomes a currency they can spend on AI. It also means the cloud providers' catalogues, quotas and regional availability start to matter as much as the model leaderboards. A model you cannot buy in your region, under your data rules, on your contract, is not really available to you, whatever its benchmark score.
Second, the operator question is now a first-class buying question. The Claude example makes this concrete: two routes, one invoice style, different data processors, different feature timing, different integration options. The same discipline applies to the OpenAI route, where Google says prompts, context and outputs stay within the customer's designated data perimeter and are not used to train foundation models, while Zero Data Retention options and Confidential Compute remain in the coming-soon category. Buyers should treat live controls and promised controls as different things, and should write down, before signing, who processes their data, where it is processed, and what happens to it afterwards. The week produced two live launches and a list of coming-soon items; confusing the two lists is the easiest expensive mistake available right now.
Third, the agent race is moving from the model to the plumbing. Claude Duet, OpenAI's Dots, Meta's Muse and Google's Gemini agent all made the same bet this month: the agent that messages you first, with a morning brief and background checks across your connected apps, wins the daily habit. That bet depends less on which model writes the best paragraph and more on connectors, permissions, memory, approvals and standards. The Agent Skills donation and the Decisions API both fit that picture: one standardises how agents learn jobs, the other standardises how agents answer narrow questions. The labs are still competing on models, but the contest that decides daily use is being fought in connectors and conventions, where switching costs are built one integration at a time.
There is a counterweight to all of this, and it arrived the same week from an unexpected direction. Replit, the coding platform, spent mid-October clarifying its own position: chief executive Amjad Masad first said at TechCrunch Disrupt that Replit was no longer working with Anthropic, then clarified on X that Replit remains model agnostic, with OpenAI as its preferred partner and default, while Claude stays selectable. Replit added Claude Haiku 5.5 a day later. The episode is a useful warning against reading any single partnership as a permanent allegiance. Defaults matter enormously, because most users keep them, and a default switch can move large volumes of spending. But defaults are also reversible, and platforms that keep rivals on the menu keep their own bargaining power. Expect more of these clarifications, not fewer, as the distribution war heats up.
What to watch next
Five dates and signals will show whether this week was a turning point or a press-release cluster. First, OpenAI's general availability date for Google Cloud, promised for the coming weeks: a firm date, published prices and regional availability will tell buyers whether the preview was a real product or a controlled demonstration. Second, enterprise adoption numbers: which large customers actually route spending through these marketplace routes, and whether their cloud commitments genuinely absorb AI bills. Third, Anthropic's expected IPO paperwork, widely reported as imminent in mid-to-late October: a public listing would put the first hard numbers on the agent-platform business that this week's announcements are selling. Fourth, the Agent Skills library under foundation governance: does the pace of new skills hold, and do rival platforms keep contributing? Fifth, the DeepSeek exclusion order's practical reach: how federal agencies, and the vendors who sell to them, treat third-party products built with DeepSeek components.
For smaller teams and individual builders, the advice is narrower and more practical. Do not migrate on announcement day. Check who operates the service you are buying, not just whose logo is on the invoice. Confirm preview prices in writing before budgeting against list prices from another channel. Sort discounts before first use, not after. And if you build agents, invest in skills and connectors that you can carry between platforms, because the one certainty in this week's news is that the platforms themselves will keep rearranging.
Frequently asked questions
Can I use OpenAI models on Google Cloud today?
Only if you are a selected enterprise customer in the private preview. Everyone else can submit an interest form, but that does not guarantee access. General availability is planned for the coming weeks, according to OpenAI, and Google has not published preview prices or a firm date.
What is the difference between Anthropic-managed and Google-managed Claude on Google Cloud?
The Anthropic-managed route is operated by Anthropic, which acts as the data processor and ships new platform features on launch day. The Google-managed route is operated by Google, which acts as the data processor, and suits customers with strict regional residency needs or native integrations such as Model Armor or Agent Engine. Both bill through Google Cloud and count toward commitments.
Does buying through Google Cloud change what the model can do?
No, the models themselves are the same families: GPT-6 Astra, Sol and Luna on the OpenAI side, and Claude Fable 5.1, Opus 5.5, Sonnet 5.5 and Haiku 5.5 on the Anthropic-managed Claude side. What changes is the purchase route, the invoice, the identity system, the commitment arithmetic and, for Claude, who processes your data.
Why would a company buy AI models through a cloud marketplace at all?
Because it can pay on one invoice, draw down an existing cloud spending commitment, and use an identity and security setup it already has. For large buyers, that removes a separate contract, a separate security review and a separate budget line. For smaller buyers, the main benefit is simpler administration rather than a price advantage.
What are Agent Skills, and why does their donation matter?
Agent Skills are reusable instructions that teach an AI agent how to perform a specific job. Anthropic donated the standard to the Agentic AI Foundation, a Linux Foundation-directed body, on October 22. If the standard stays open and widely used, skills built for one platform can travel further, which lowers switching costs for agent builders.
Should small businesses switch their AI provider because of this news?
Probably not yet. The OpenAI route is still in private preview, and the Claude routes differ mainly on data-processing and feature timing, which matter most at enterprise scale. Small teams gain more from testing one model on one real workload, and from keeping their prompts, skills and connectors portable, than from chasing launch-week announcements.
The bottom line
This was not a week about a smarter chatbot. It was a week about who sells intelligence, on whose invoice, under whose data rules, and through whose standards. OpenAI and Claude arriving on Google Cloud within a day of each other does not settle which lab builds the best model. It settles something else: frontier AI is now a product you buy through a shop you already use, and the shop's terms, commitments and identity systems are part of the product. The labs still matter. The plumbing now matters as much. Buyers who understand both will spend less and switch more easily than buyers who only follow the leaderboards.
For more on technology, markets and the business of AI, visit abdulrahimmughal.com.