🎧 Prefer to listen?
OpenAI’s models are moving into Google Cloud — Model Garden and Antigravity, Google’s agent-powered coding tool included — which quietly completes a shift most solo builders haven’t noticed yet: you can now reach every frontier model family through a single cloud relationship. Google’s Model Garden already hosted Gemini, Claude, and 200-plus open models; adding the GPT 6 and 6.1 family (Astra, Sol, Luna) makes Google the only large cloud offering all three leading closed LLMs. When I read the announcement, I didn’t think about enterprise procurement. I thought about my own bills — and spent 20 minutes mapping which of my existing commitments already unlock which models. The answer changed two things about how I’ll buy AI this year.
The deal itself is straightforward: OpenAI’s models run entirely on Google’s infrastructure, billed through the Google Cloud Marketplace, with Microsoft still OpenAI’s primary partner. The interesting part for you is what Google built around it — the ability to dynamically move workloads between LLMs based on performance, latency, and efficiency. That’s not just hosting; that’s routing. The industry is converging on a model where which lab made your model matters less than which platform can route your tasks to the right one. This is the same consolidation pressure behind Google’s internal AI shake-up and the pricing pressure behind what AI subscriptions are actually worth — and the $100/month Google AI Ultra plan now has enterprise-grade backup for its positioning.
What I actually did
I listed every AI spend commitment my business carries: the consumer subscriptions, the API accounts, the annual plans, the cloud credits. Then I marked which models each commitment gives me access to — and, crucially, which of those access paths I was actually using.
The finding: I was paying for model access I never touched. One annual subscription included a frontier model tier I’d never once used, while I’d been considering a new subscription to reach a model that an existing commitment already covered. That’s the solo-builder version of the enterprise problem this deal solves — fragmented access, invisible overlap, spending that doesn’t match usage. For teams, the deal means GPT models can arrive inside existing Google Cloud commitments; for you, the equivalent question is what your existing commitments already unlock.
Do this yourself
Twenty minutes, no code:
- List every AI commitment. Subscriptions, API accounts, credits, annual plans — anything recurring that buys you model access. Include the ones you forgot about; that’s usually where the finding is.
- Map models to commitments. For each, note which model families you can reach: GPT, Claude, Gemini, open models. This is your actual access map — it’s usually bigger than you think.
- Find the overlap. Two commitments reaching the same model family? That’s a cancellation candidate. One commitment reaching a family you use elsewhere? That’s a consolidation opportunity.
- Check the routing angle. If your workflow tools let you swap models per task — most agentic tools now do — configure the task-appropriate model instead of the vendor-loyal one. Writing tasks, coding tasks, and analysis tasks often have different best-fit models, and model fallback planning works better when the fallback is a different family, not just a different version.
- Re-price the stack. With your true access map in hand, price the minimum set of commitments that covers every model family you actually use. The answer is often one or two fewer bills than you’re currently paying.
What failed
Two honest misses.
I initially confused access with quality. Having GPT models available through Google Cloud doesn’t mean the experience matches OpenAI’s own platform — features, latency, and tooling can differ by distribution channel. My test of one alternative path worked for generation but lacked the features I rely on. Access is not parity; verify the workflow, not just the model name.
The “dynamic switching” future isn’t here for solo tools yet. Google’s workload routing is an enterprise platform feature. At the solo scale, you’re still choosing models manually — which is exactly why the access map matters now: when routing does reach consumer tools, you’ll want your commitments already consolidated.
Who should skip this
If you use one consumer AI subscription and it covers everything you need, skip this — the deal changes enterprise procurement first, and consumer plans on a slower timeline.
Skip it too if you’re deliberately vendor-loyal and happy; simplicity has real value. But if you pay for multiple AI access paths, run automations across model families, or have cloud credits sitting half-used, this announcement is the nudge to audit what you already own. And as the labs keep racing on capability and price — with custom chips reshaping inference economics underneath — the builders who win are the ones whose access is deliberate rather than accidental.
The takeaway
The walls between AI labs are coming down at the distribution layer: one cloud, three frontier families, routing that picks the right model per task. The enterprise version of this convenience is already live. Your version costs twenty minutes with a spreadsheet — list your commitments, map your access, cancel the overlap. The models are converging; your bills should too.