On September 1, Anthropic shipped Claude Fable 5.1 and Mythos 5.1 at unchanged list pricing with three breaking API changes. On September 2, Google shipped Gemini 3.8 Flash and printed an end date on its introductory price. Two weeks later a survey of a hundred enterprise CIOs found that 37 percent are running five or more AI models in production, up from 29 percent the year before. Eighty-one percent said they were concerned about depending on a single AI vendor. Forty-seven percent said losing their primary AI vendor would break a business function. Six percent said they could switch cleanly.
That last number is the one to sit with. It says most of the market has locked themselves into a vendor whose product keeps changing under them and whose price keeps moving on a schedule they do not set.
Choosing between AI vendors, as most executives frame it, is a decision about which lab wrote the best model this quarter. The frame is a trap. The lab that wrote the best model last quarter almost never wrote the best model this one, and the buyer who priced a two-year plan on the leaderboard has already lost the plan. The choice worth making is architectural. It is about which parts of your business you keep on your side of the API and which parts you rent from whichever supplier is currently ahead.
The Microsoft lesson
Read what happened between Microsoft and OpenAI on April 27, 2026 as a lesson about what a vendor lock actually means. The two companies restructured their partnership. Azure exclusivity ended. OpenAI can now serve every product to customers on AWS, Google Cloud, Oracle, or any infrastructure partner it prefers. Microsoft holds a license to OpenAI IP through 2032, and the license is no longer exclusive. Microsoft stopped paying the revenue share. The clause that would have re-shaped Microsoft's contractual rights once OpenAI declared AGI was cut entirely.
Read the amendment through the buyer's eye. The most consequential contract in enterprise AI, the one that founded Azure OpenAI Service and gave Microsoft the moat it built Copilot on, was rewritten in place. Every AI vendor contract you sign in 2026 is being signed under this precedent. The counterparty on the other side of your deal watches Microsoft, notices that a two-hundred-fifty-billion-dollar strategic pillar can be renegotiated in place, and understands that yours can be renegotiated too.
Buyers who read the amendment as good news, more optionality, more clouds, missed the second-order point. If Microsoft could not lock in a supplier it invested thirteen billion dollars into, you cannot lock in a supplier by writing a five-year enterprise agreement. The lock has to sit somewhere else. Either you build the lock yourself, in the shape of an architecture that keeps the vendor swappable, or you accept that a supplier gets to rewrite the terms of your relationship whenever their frontier shifts.
Anthropic runs the same model on three clouds
Anthropic ships Claude on AWS Bedrock, Google Cloud Vertex AI, and Azure Foundry. It is the only frontier AI model currently available on all three. On May 1 the company signed for up to five gigawatts of AWS Trainium and Trainium3 capacity over ten years, committing more than a hundred billion dollars. On October 23, 2025 the company signed for up to a million Google TPUs across a partnership that brings over a gigawatt of Google compute online in 2026 alone. Anthropic's run-rate revenue crossed thirty billion dollars in 2026, up from about nine billion at the end of 2025, with over a thousand business customers each spending more than a million dollars a year.
Look at that shape. Anthropic is spending nine figures a quarter on capacity contracts with three infrastructure providers who are supposed to be locking each other's customers in. It sits on top of AWS's silicon, Google's silicon, and NVIDIA's silicon at the same time. Its list of frontier peers is short and every peer has picked exactly one bet.
Anthropic's answer to the vendor-choice question, applied to its own supply chain, is to have more than one supplier and to make each of them believe that being the only one would be worse for them than sharing. The customer sitting on top of Anthropic inherits some of that discipline, but only if the customer's integration is written against the seam that Anthropic exposes and not against a proprietary abstraction one cloud provider layered on top.
Six percent
Here is the number I cannot stop thinking about. Six percent of enterprise executives say their company could switch AI vendors without material disruption. Ninety-four percent cannot. Half of them would break a business function. The rest would take a hit somewhere else in the P&L.
The number was chosen. Nobody sat down and picked it. Every workflow written against a vendor's tool schema chose a fraction of it. Every eval set graded against a specific model's failure modes chose a fraction of it. Every agent that memorized a prompt structure a vendor blessed one week, and quietly deprecated on the next major release, chose a fraction of it. Multiply by three years of accumulated integration and you get the aggregate lock-in of an enterprise stack.
The six percent who could switch did something different. I have worked inside a handful of them. The pattern is unfailingly the same. They wrote the router themselves. They wrote the eval harness themselves. They kept the memory system on their side of the API. They shaped their tool schemas to a form they compiled down to each vendor's grammar rather than adopting a vendor's grammar as their own. In every case the model call was the smallest piece of code they shipped, dwarfed by the seam around it.
The seams that keep the choice open
Four seams determine whether the vendor choice is reversible. Skip any one of them and the six percent bracket closes to you.
The first is the routing seam. The router decides which model gets which call. At 2026 prices, a company routing carelessly to a single frontier vendor pays two to five times what a routed stack pays. More importantly, the router is the only place in the system where the model identity is a variable. Every other layer should treat the model as a function of the router's output.
The second is the evaluation seam. Every AI system in production is graded against a set of traces. If the traces live inside a vendor's console, the vendor owns the memory of what worked. Move the traces to your own store. Grade against a rubric your team wrote. When Anthropic ships 5.1 or Google ships 3.8 Flash, feed the same eval through the new model. This is what the six percent do the week a new frontier ships. The rest of the market waits for the vendor's benchmark, notices six months later that their behavior changed, and starts a re-tuning project from scratch.
The third is the memory seam. State is the vendor's favorite lock. Every conversation history, every embedding, every retrieval index that lives in a vendor's own store becomes a vendor artifact. Buyers who ran Pinecone in 2023, migrated to a hyperscaler's vector store in 2024, and now run against a vendor-native memory service in 2026 have paid the memory tax three times without noticing it was a tax. Keep memory in a store you own. Ship embeddings across API boundaries yourself.
The fourth is the tool seam. Every model has a tool-calling grammar. OpenAI has Responses. Anthropic has tool_use. Google has function declarations. xAI has its own. If your business logic speaks any single one of these directly, your business logic just became that vendor's dependent. Write a tool schema of your own, expressive enough to cover what your business does, and compile it down into each vendor's grammar. Every one of the six percent has this.
None of these four is glamorous. None of them is what a systems integrator will sell you as a project because the value of the seam is precisely that it stays out of the way. Owning the seams is a discipline. Owning them is the only path to being the six percent.
The two-vendor rule
The single most useful heuristic I give to a CEO who is choosing between AI vendors is this. Whoever you pick, keep a second vendor warm at all times. Warm means a working prototype of every workflow routed through the second vendor's model at any given moment, graded against the same eval set, connected to the same memory store, calling the same tools. It does not mean an enterprise commitment to the second vendor. It does not mean a full production stack behind it.
The second vendor costs you between five and fifteen percent of your top-line AI budget in most stacks. It is the cheapest strategic asset available on the AI cost curve, and it is what turns you from a hostage into a customer. When your primary raises prices, ships a breaking API change on a Sunday night, or quietly deprecates a model line, the second vendor becomes your production path in a week. You do not negotiate a discount by mentioning the second vendor. You negotiate a discount by pointing at last quarter's traffic mix.
The two-vendor rule does one more thing. It forces you to keep the seam clean. A team that runs everything on one vendor rationalizes vendor-shaped code and stops noticing where the seam has broken. A team that runs a warm second vendor cannot rationalize it. Every unclean seam breaks the second vendor's build. The build tells you where the lock has quietly grown back.
Picking the vendor
Given the seam is right and the second vendor is warm, the picking becomes a much shorter exercise than most enterprises make it. Three questions.
Which vendor has the model closest to your workload's edge, right now, at your volume. Grade against the top of your workload at your volume. A leaderboard rank cannot make this call. If you handle fifty thousand small-context high-frequency tool calls a day, the cheapest capable frontier model with structured output wins. If you handle two hundred long-context reasoning tasks a day, the frontier reasoning model wins even at ten dollars per million input tokens and fifty per million output. If you handle a compliance-sensitive workload, the vendor whose regional deployments cover your buyer's jurisdictions wins.
Which vendor has the deployment path that fits inside your existing cloud commit. If you have three years left on an AWS enterprise discount, running Claude on Bedrock or GPT-6 on the AWS deployment routes at a real dollar discount. If you have a Google Cloud commit tied to a data warehouse, Gemini in Vertex clears at a price no hyperscaler-neutral vendor can beat.
Which vendor will still be improving the model on your workload two years from now. This is the harder question. Anthropic's revenue trajectory says it can afford to keep training. Google can afford to keep training. OpenAI, now free to sell everywhere, has the flexibility it needs. Smaller labs will get squeezed. Bet your primary on a lab whose economics can survive a downturn. Bet your second on a lab that gives you a different failure mode.
Do this and the choice takes an afternoon. Most companies spend six months on it because they have not built the seam and are using the vendor selection to make an architecture decision through the back door. That sequence costs them the six percent bracket. It costs them their next contract renegotiation. It costs them the surface a systems integrator will later charge to peel back off the stack in 2028.
What choosing looks like when the seam is right
The 37 percent CIO figure hides a distribution. Some of those companies run five vendors because their teams could not agree on one. Others run five because their eval harness picked five for five different workload shapes. The second group is the group that clears the six percent bar.
The pattern in that second group is unfailingly the same. Their router logs show a mixed distribution across two or three frontier vendors and a fine-tuned open-weight model on their own hardware for the highest-volume shapes. When Anthropic shipped 5.1 on September 1, that group ran the eval harness the same afternoon. The workflows that improved moved to 5.1 through a router config change. The workflows that regressed stayed on 5.0 for the week and got re-tuned. When Google shipped 3.8 Flash on September 2, that group moved a fraction of its traffic to the cheaper model and saved money by lunch. Nothing broke. Nothing needed a project.
The group that cannot do this is the 47 percent whose primary vendor loss would break a business function. Their eval harness lives in a vendor console. Their memory lives in a vendor's vector store. Their tool schemas match a vendor's grammar. Every model release from their primary vendor is an event they must respond to, and every model release from a competing vendor is an event they cannot respond to without a project.
That is what buying where you can leave looks like. No vendor sells it as a package. A vendor cannot sell you the ability to walk away from that vendor.
The imperative
Every executive team is going to make a vendor choice this quarter. Most of them will pick the vendor first and worry about the architecture later. That sequence produces the ninety-four percent number. It produces the six-month re-tuning project every time a supplier ships. It produces the negotiation where the vendor holds every card.
Architect the seam first. Pick two vendors, not one. Keep the second warm. Move the eval, the memory, the router, and the tool schema onto your side of the boundary. Make the model call the smallest surface in your stack. That is the work. Nobody sells it to you as a package because a package makes you a customer of the packager.
Agor AI Advisory builds this architecture with executive teams before the enterprise agreement gets signed, so the agreement is signed from strength. We write the router. We stand up the eval harness. We move the memory to your side. We define the tool schema your business speaks. We keep the second vendor warm. We do this in weeks, not quarters, and we do it once so that the next model release from any vendor is a config change rather than a project.
The choice worth making is not which vendor you buy from. It is where you draw the line between your business and any vendor.
Sources
- The next phase of the Microsoft OpenAI partnership, OpenAI, April 27, 2026
- The next phase of the Microsoft-OpenAI partnership, Microsoft, April 27, 2026
- Anthropic and Amazon expand compute collaboration, Anthropic, 2026
- Anthropic expands Google and Broadcom compute deal, Anthropic, 2026
- From Lock-In to Leverage: How the OpenAI-Microsoft Partnership Amendment Redefines Enterprise Multi-Cloud AI Procurement in 2026, FifthRow
- Avoid AI Vendor Lock-In: A Multi Model AI Strategy 2026, AvePoint
- AI Model Releases: September 2026 Tracker and Dated Ledger, Digital Applied