OpenAI puts GPT-6 on Google Cloud: 3 models, 1 invoice, and $0.10 per million cached tokens

OpenAI will sell its GPT-6 models and Codex coding workflows through Google Cloud’s Model Garden, the company announced on Oct. 21 — putting a direct rival to Google’s own Gemini on Google’s enterprise storefront. Cached GPT-6.1 Sol input costs $0.10 per million tokens, one-twentieth the $2 standard rate.

Article Brief

Key Takeaways

6 Points36s Read

• 01The deal–OpenAI’s frontier models and Codex agentic workflows are coming to Google Cloud’s Model Garden, announced Oct 21. Private preview is live now; general availability is planned in the coming weeks.
• 02The lineup–GPT-6 Astra (flagship), GPT-6.1 Sol (near-Astra intelligence at lower cost) and GPT-6 Luna (fast, cost-efficient). Codex CLI and desktop workflows can already point at Google’s Gemini platform for inference; Google Antigravity and the Gemini Agent follow.
• 03The money mechanics–Purchases route through Google Cloud Marketplace onto a single Google Cloud invoice. Enterprises can draw down existing GCP spend commitments against OpenAI inference, and usage counts toward Flexible Savings Plans.
• 04TECHi’s price math–OpenAI’s published rates put GPT-6.1 Sol at $2/$0.10/$10 per million input/cached-input/output tokens; Astra implies $10/$50 (Sol is one-fifth of Astra, and Ultrafast at 1.2x Astra costs $12/$60). Cached input is one-twentieth of standard input.
• 05The strategy–Google Cloud claims it is now the only hyperscale cloud with all three leading closed LLMs — Gemini, Claude and GPT — on one platform. OpenAI says Microsoft remains its primary cloud partner.
• 06What it isn’t–No Google-channel pricing, regions or GA date disclosed. Zero Data Retention and Confidential Compute for the OpenAI workloads are both ‘coming soon.’

A rival’s models, on Google’s own cloud

The announcement came in OpenAI’s official post: its frontier models and Codex agentic workflows are coming to Google Cloud’s Model Garden, with general availability planned “in the coming weeks.” Google’s own announcement, published the same day, confirms the models are rolling out now in private preview to select enterprise customers.

The strategic weight of the sentence is hard to overstate. For most of the ChatGPT era, OpenAI’s models were effectively a Microsoft Azure exclusive on the hyperscale side — OpenAI’s own post takes care to note that “Microsoft remains OpenAI’s primary cloud partner,” a line that reads as both reassurance and epitaph. The company is now a multi-cloud vendor, and the customer relationship runs through whoever owns the enterprise’s cloud contract.

That is where Google’s mechanics matter. Google Cloud customers will buy OpenAI models through the Google Cloud Marketplace, with usage appearing on a single Google Cloud invoice. Enterprises with Google Cloud spend commitments can draw those down against OpenAI inference, and usage counts toward Google’s Flexible Savings Plans. In procurement terms, OpenAI just became a line item inside a budget that already exists — no second vendor contract, no separate procurement cycle.

The three-model ladder — and TECHi’s price math

Google names the opening lineup: GPT-6 Astra, the flagship for demanding work across coding, computer use, professional tasks and scientific research; GPT-6.1 Sol, near-Astra intelligence at lower cost for agentic coding and professional work; and GPT-6 Luna, the fast, cost-efficient model for everyday work at scale. Customers in the preview can call the models through Google’s APIs today, and can point their local Codex command-line tool and Codex desktop app directly at Google’s Gemini platform for inference. OpenAI’s models are also slated for Google’s Antigravity coding platform and the Gemini Agent.

The GPT-6 lineup on Google Cloud

Model	Positioned for	Google’s description
GPT-6 Astra	Coding, computer use, professional tasks, scientific research	The flagship, for demanding work
GPT-6.1 Sol	Agentic coding, computer use, professional work	Near-Astra intelligence at lower cost
GPT-6 Luna	Everyday work at scale	Fast, cost-efficient

The pricing tells the real story of the ladder, and it can be reconstructed from OpenAI’s own published rates. OpenAI has said GPT-6.1 Sol’s Ultrafast tier costs $12 per million input tokens and $60 per million output tokens — “just 1.2x the cost of Astra.” Divide through and Astra’s standard pricing implies $10 per million input tokens and $50 per million output tokens. Sol’s standard tier runs $2 per million input tokens, $0.10 per million cached input tokens, and $10 per million output tokens.

TECHi’s price math: the GPT-6 ladder

Tier	Per 1M input tokens	Per 1M output tokens	TECHi read
GPT-6 Astra (standard, implied)	$10	$50	5x Sol’s rates; implied from OpenAI’s stated ratios (Sol = one-fifth of Astra; Ultrafast = 1.2x Astra)
GPT-6.1 Sol — Ultrafast	$12	$60	1.2x Astra; for speed-critical agentic work
GPT-6.1 Sol — standard	$2 ($0.10 cached)	$10	Cached input is 1/20 of standard input — the real price of agentic coding

Put plainly: Sol’s output costs one-fifth of Astra’s ($10 vs. $50), and cached input at $0.10 per million tokens costs one-twentieth of Sol’s own standard input rate ($2). For agentic coding workloads — where the same repository context gets replayed across many turns — the effective price of intelligence collapses toward the caching rate, not the headline rate. That is the number enterprise buyers will actually pay attention to, and it is why Google’s single-invoice, commitment-drawdown packaging matters: the marginal dollar of OpenAI inference can now come out of cloud spend the CFO already approved.

Codex is the wedge, not the catalog

The model listings get the headlines, but Codex may be the sharper edge of the deal. Developers already run Codex in their terminals and IDEs; Google says private-preview customers can today connect their local Codex CLI and desktop workflows directly to Google’s Gemini platform for inference. The developer doesn’t change tools, doesn’t change workflow, doesn’t sign a new contract — the inference just routes through the cloud they already pay for.

This is the pattern that made cloud marketplaces the quiet winners of the AI boom. The buyer was never going to issue a purchase order for “tokens.” But “draw down your existing GCP commitment, see it on the same invoice, and have it count toward your Flexible Savings Plans” is a sentence a procurement department can say yes to in an afternoon. OpenAI’s path from evaluation to broader adoption, in its own words, now runs through Google’s billing system.

It also deepens an odd but durable dynamic: the agent era is being built on infrastructure owned by companies whose own models compete with the agents running on top. Google will sell you Gemini, Claude, or GPT — and take the infrastructure margin either way.

The cloud scoreboard, redrawn

Step back and the three-hyperscaler map looks different than it did a year ago. OpenAI’s models now have a commercial route on Azure, AWS and Google Cloud — Microsoft’s one-time lock on OpenAI distribution is over, even as OpenAI keeps calling Microsoft its primary cloud partner. Google’s claim to be the only hyperscaler with all three leading closed models on one platform is the kind of line that forces a response: AWS and Azure cannot let the “one platform, every model” pitch stand unchallenged for long.

The counter-argument writes itself, of course. Distribution is not exclusivity, and a Model Garden listing is not a moat. Google’s bet is that when every cloud sells every model, the differentiator becomes governance, billing and procurement convenience — the unglamorous enterprise plumbing that decides where the budget actually goes. The next two earnings seasons will show whether enterprises agree.

Why Google wants this, and why OpenAI agreed

Google’s blog makes the boldest claim of either announcement: Google Cloud is now “the only hyperscale cloud provider offering access to all three leading closed source LLMs” — Gemini, Claude and GPT — alongside more than 200 first-party, third-party and open models on one platform. If true, it is a genuine differentiator against Azure and AWS: the enterprise buyer who wants model choice without multi-cloud sprawl has exactly one place to get it.

The pitch is wrapped in governance language that regulated industries will read closely. Google says customer prompts, context and outputs stay entirely within the customer’s designated data perimeter and are never used to train foundation models, with Zero Data Retention options “coming soon.” OpenAI workloads run on dedicated Google Cloud infrastructure, queries route through hardened, isolated gateways, and every agent and model interaction operates under an authenticated identity with unified audit logging. The subtext: this is the version of OpenAI built for banks, hospitals and governments.

For OpenAI, the logic is distribution. Its post frames the move as giving enterprises “greater choice in how they discover and use frontier models, with a path from evaluation to broader adoption” — the funnel language of a company that has realized the fastest way to reach enterprise budgets is through the cloud vendor that already invoices them. The irony, which nobody at either company will state, is that Google is now hosting and billing for the models that compete most directly with its own Gemini line — while its cloud chief and OpenAI’s president both publicly bless the arrangement.

What to watch

• General availability timing. “The coming weeks” is doing a lot of work — watch whether the GA date lands before Microsoft’s Oct. 28 earnings, when Azure’s AI growth gets its own spotlight.
• Google-channel pricing. Neither company has said what the models will cost through Google Cloud relative to OpenAI’s own API rates, or which countries and regions get access first.
• Antigravity and the Gemini Agent. OpenAI models arriving inside Google’s own coding surfaces will be the real test of the “model choice flexibility” pitch — and of how comfortable Google is surfacing a rival inside its own products.
• The commitment drawdown in practice. The first earnings season after GA should show whether enterprises actually redirect GCP spend commitments toward OpenAI inference, and whether that cannibalizes Gemini usage or expands the total.
• Microsoft’s response. “Primary cloud partner” is a title, not a contract. Watch what Azure announces to match the single-invoice, all-models-in-one-place pitch.