OpenAI puts Astra, Sol and Luna in Google Cloud, billable against GCP commitments
On October 21 OpenAI put GPT-6 Astra, GPT-6.1 Sol and GPT-6 Luna into Google Cloud's model garden in private preview, billable against GCP commitments.
On October 21, OpenAI and Google Cloud said GPT-6 Astra, GPT-6.1 Sol and GPT-6 Luna are now listed in Google Cloud’s Model Garden, the catalog where Google sells its own, third-party and open models side by side. The first wave is a private preview for selected enterprise customers. What is new is the billing: all three models are transacted through Google Cloud Marketplace and draw down cloud commitments the customer already signed. Google adds that the usage lands on a single Google Cloud invoice.
Key points
- Private preview, not general availability. The three models are rolling out to “select enterprise customers.” OpenAI’s post says general availability is “planned in the coming weeks” and gives no date.
- Google’s marketplace, Google’s commitments. Marketplace transaction and drawdown against existing commitments appear in both companies’ posts. The single-invoice line and counting toward Flexible Savings Plans appear only in Google’s blog post.
- Developer access starts today, narrowly. Local Codex CLI and Codex Desktop workflows can connect to the Gemini platform for inference now. Antigravity and the Gemini Agent come later; Google’s post calls the latter “Gemini for business.”
Background
Microsoft’s exclusive license on OpenAI’s intellectual property ended this year. According to CNBC, Amazon announced plans to host OpenAI models the day after the exclusivity ended and started doing so in April. At its September 29 DevDay, OpenAI put managed agents on AWS and moved ChatGPT into Slack and Teams, which we covered at the time. CNBC also notes an earlier partnership: OpenAI was already using Google Cloud to run ChatGPT and its API. That was buying compute. This is putting models on a shelf, which is a different transaction.
Google spent the past month clearing the shelf. Gemini 4 Argon shipped on September 30 and opened to everyone on October 13 (our entry), and Antigravity already carries models from both Gemini and Claude. With GPT added, Google’s blog makes a sharp claim: Google Cloud is the only hyperscale cloud that offers all three leading closed-source frontier model families. The claim is Google’s alone; Ars Technica attributes it as “Google says.” The same Ars piece did add a qualifier, “at least if you ask Google,” to a different claim: that Gemini 4 Argon has caught up with the competition.
The facts
Two official posts and four newsrooms agree on the following. Note what is missing: no price list for the Google channel.
| Item | Detail |
|---|---|
| Models | GPT-6 Astra (flagship), GPT-6.1 Sol (near-Astra, cheaper), GPT-6 Luna (fast, low cost, high volume) |
| Status | Private preview for select enterprise customers; GA “in the coming weeks” |
| Transaction | Google Cloud Marketplace, draws down existing commitments; single invoice per Google only |
| Discount path | Counts toward Flexible Savings Plans (Google’s claim only) |
| Developer access | API builds available today; local Codex CLI and Codex Desktop route inference to the Gemini platform |
| Coming next | Google Antigravity and Gemini Agent (Google’s post says “Gemini for business”) |
| Governance | Data stays in the designated perimeter and is not used to train foundation models; ZDR and Confidential Compute both marked “coming soon”; VPC Service Controls and unified audit logging |
| OpenAI list price, per 1M tokens | Astra $10 in, $50 out; Sol $2 / $10; Luna $0.10 / $0.50 (pricing page read October 22, 2026) |
Governance gets the most space in Google’s post, and it is the part readers skim. OpenAI models run on “dedicated Google Cloud infrastructure,” every query passes through isolated gateways, third-party model providers have no unauthorized access to the customer’s cloud environment, and each call runs under an authenticated identity that IT admins can trace. Two pieces of that stack are not live: zero data retention and Confidential Compute are both “coming soon.” SiliconANGLE adds that the Confidential Compute step will isolate part of a server’s memory with trusted execution hardware so only one workload can read it.
The discount behind the Flexible Savings Plan line sits in a separate Google post from August 26: a one-year commitment takes 10% off, a three-year commitment 20%, with no minimum or maximum monthly spend. That post described Gemini Enterprise usage; the October 21 post links to it and says OpenAI model usage counts toward the same plans. How much a buyer saves depends on the term it already signed.
On price, Google and OpenAI published no Google Cloud SKU or rate. The only comparison available is OpenAI’s own list price, where Sol costs one fifth of Astra and Luna one hundredth. That arithmetic is where SiliconANGLE’s “five times” and “100 times” figures come from; it is not a new price sheet. Neither company says whether the models run on TPUs or Nvidia GPUs inside Google’s data centers.
What others say
Google Cloud CEO Thomas Kurian told SiliconANGLE: “We are focused on giving customers a broad selection of models, including those from OpenAI.” He framed the point as picking the right model for each job and balancing quality against cost, which is the same frame Google’s blog uses: swap models by performance, latency and cost without signing a contract per vendor.
The customer voice in Google’s post is Yashodha Bhavnani, VP of AI products at Box, who said access to OpenAI’s frontier models alongside Google’s security and compliance work lets Box get more value from enterprise content. It is a standard early-customer statement with no usage or performance numbers attached.
Ryan Whitwam at Ars Technica picked out two loose ends. Google has not confirmed that Antigravity will get these same three models. And Argon “burns through tokens at an astonishing rate,” which gives Antigravity users a reason to want an alternative in the picker. He also did not let the awkward line slide: OpenAI’s post ends by stating that Microsoft remains its primary cloud partner.
CNBC’s Jordan Novet supplied the business frame. Google Cloud’s annualized revenue growth accelerated to about 82% in the second quarter from 63% in the first, against 37% at Amazon and 42% at Microsoft, and all three report earnings the following week. He also notes OpenAI is raising at a $1.4 trillion valuation and aiming for an IPO. Those numbers come from CNBC alone; neither company’s post mentions them.
Hacker News picked the story up on October 21 (250 points and 107 comments when read on October 22). The leading thread mocked the “Microsoft remains our primary cloud partner” sentence, with one reply calling it something you say because a contract requires it. A second thread explained the billing better than either vendor did: an enterprise that committed to annual GCP spend on compute and storage can now move idle budget onto model calls instead of losing it. Someone asked whether Google might distill OpenAI’s models while hosting them; the answers rested on contracts and audit logs, and nobody offered a technical guarantee.
Our take
For a team that already holds an annual Google Cloud commitment, this changes the procurement path. Model calls move from “a separate contract and a separate budget line” to “a slice of a commitment we already have.” That is true for any cloud customer, but it matters most for the ones who chronically underspend their commitment, because converting leftover budget into GPT tokens beats buying compute nobody uses at year end.
The caveats are long. Google has not published admission criteria for the preview, so filling in the interest form does not mean access. ZDR and Confidential Compute are both future tense, which leaves regulated industries holding a blog promise for now. The Flexible Savings Plan eligibility is Google’s claim, missing from OpenAI’s post. Price, regions and rate limits for the Google channel are unpublished, so budgeting from OpenAI’s list price will probably not match what Google bills.
The one thing a developer can do today is point local Codex CLI or Codex Desktop at the Gemini platform. Google says “starting today,” but published no command, environment variable or endpoint documentation, so trying it means waiting for docs or asking an account team. Antigravity users wait for Google to sort out tiers: that product currently runs Gemini and Claude, with Claude Opus and Sonnet 4.6 open to free, Plus and trial accounts until November 2, while Opus and Sonnet 5.5 need a paid Pro tier or higher (9to5Google).