Stop Picking a Model. Pick a Control Plane.

I sat in a vendor bake-off recently where two smart teams spent three hours arguing whether Gemini 3.1 Pro or the latest Claude was the better enterprise model. It was a good debate. It was also the wrong meeting. By the time their procurement cycle closes, the model leaderboard will have flipped at least once, and — here's the part everyone missed — both models are available on both platforms they were choosing between anyway.

The enterprise AI decision in 2026 has quietly stopped being about which model. It's about which control plane you standardize on. And that's a much more durable question, because the control plane is the thing you'll actually live inside for the next three years: the governance, the policy enforcement, the cost controls, the orchestration. The model is a dropdown.

The platforms converged on multi-model

Look at what the two big clouds actually shipped. Microsoft renamed Azure AI Foundry to Microsoft Foundry — dropping "Azure" specifically to signal that agents are first-class citizens, not just another cloud service (Medium / Azure in Practice). More importantly, Foundry is now an explicitly multi-model control plane with OpenAI, Anthropic, Mistral, DeepSeek, and Microsoft's own MAI models on one platform, plus a catalog north of 11,000 models (EPC Group).

Google did the structurally identical thing. Vertex AI became the Gemini Enterprise Agent Platform in April 2026, and it offers first-class access to 200+ models through Model Garden — including Google's own Gemini 3.1 line and Claude, sitting side by side (Google Cloud Blog). Notably, all Vertex AI services and roadmap now flow exclusively through the Agent Platform — there's no standalone Vertex to fall back to.

So when your team debates "Gemini vs. Claude," the honest answer is: you can have both, on either platform. The model choice is real but reversible — you can swap it next quarter. The platform choice is the one with switching costs measured in re-platformed governance and retrained teams.

What you're actually choosing

If the model is a dropdown, what are you really evaluating when you pick a control plane? Four things, none of which show up in a benchmark:

Governance and policy. Foundry's pitch is consistent security, compliance, and policy controls across every agent (Microsoft). Gemini Enterprise bundles enterprise governance into the same pay-as-you-go platform as the build tools (Google Cloud). This is the part you'll lean on hardest and the part that's most painful to migrate. Evaluate it like it's the whole product, because operationally it is.

Identity and data gravity. This is usually the real tiebreaker, and it's boring on purpose. If your enterprise lives in Microsoft 365, Entra, and Purview, Foundry's integration story is a force multiplier. If your data and identity sit in Google Cloud, the Gemini platform's native ties win. Pick the control plane that's closest to where your data and your identity provider already are — you'll fight fewer integration battles and close fewer security gaps.

The agent runtime and orchestration. Both platforms now ship managed runtimes — Foundry's Agent Service with the Responses API as a single entry point, Gemini's Agent Engine with a code-first ADK and a low-code Agent Studio. If your near-term roadmap is multi-agent workflows (and for most teams it is), this layer matters more than the model. Build a real workflow on each before you sign.

Cost model. Both are usage-based now. That's good for governance and bad for surprises. Whichever you pick, you want per-team and per-workflow token visibility on day one — the platform that makes spend legible to you is worth more than the one with the cheaper headline rate.

A pragmatic way to decide

I tell teams to run the evaluation in this order, which is roughly the inverse of how the exciting meeting wants to run it:

• Start with data gravity and identity. Where does your sensitive data already live, and who's your identity provider? This usually narrows it to one obvious choice and saves you the bake-off.
• Pressure-test governance with your actual compliance requirements, not a demo. Can you enforce your real policies across every agent, centrally?
• Build one production-shaped agent workflow on the finalist and measure the orchestration and cost-visibility experience.
• Only then argue about models — and treat that argument as a quarterly decision you'll revisit, not a permanent commitment.

If you do this, the model debate shrinks to its proper size: a tuning knob you adjust as the leaderboard moves, inside a platform you chose for reasons that won't change next month.

The takeaway

The model you pick today will be outdated by your next planning cycle; the control plane you pick will outlast several model generations. Both Microsoft and Google have made the models commoditized and interchangeable on purpose — the moat they're building is the governance, identity, and orchestration layer. Spend your evaluation energy there. Choose the control plane closest to your data and identity, prove its governance against your real requirements, and let the models be the easy, reversible part. That's not settling for less. That's recognizing where the durable decision actually lives.

Sources: Microsoft Foundry rename explainer · EPC Group: Foundry multi-model control plane · Google Cloud: introducing Gemini Enterprise Agent Platform · Microsoft Foundry Agent Service overview · Gemini Enterprise Agent Platform product page