Model Lock-In Cracks as Vercel AI Agents Pick Labs

If Vercel AI agents can choose between OpenAI GPT-5.6 launch, Anthropic, Gemini, Deepseek, and GLM-5.2, who actually owns the application: the lab that made the model, or the platform that decides when the model gets used?

That is the fight underneath Guillermo Rauch’s interview with TechCrunch. Vercel now sees 6 million deployments a day, half triggered by coding agents, and more than 1 trillion tokens pass through its AI gateway daily. Those numbers explain why Rauch is talking less like a cloud infrastructure CEO and more like someone trying to define the control layer for production AI.

Can Vercel AI agents break model lock-in before the labs lock up the agent layer?

Rauch’s core argument is simple: prototypes tolerate model loyalty. Production systems don’t.

Last year, in his telling, teams were still in the “unleash the agents” phase. Vercel itself had “hundreds of agents organically developed and deployed within the company.” Then the practical problems arrived. Security. Audit trails. Tool calls. Access controls. Data leaving places it shouldn’t.

That shift matters because Vercel AI agents sit in the part of the stack where decisions get made. The model generates. The agent acts. The platform controls what the agent can access, where it runs, which tools it calls, and potentially which model gets the job.

Rauch frames the issue as a split between the model and the agent:

“I really think at this point we’re deciding on whether the model and the agent are going to be coupled. Do you get all your intelligence from one place? Or do you get a module or a library or a building block from one provider, and then you build on top of it.”

XOOMAR analysis: this is the real strategic claim. If agents remain tightly coupled to one model provider, labs betting on agents keep the center of gravity, a dependency tension also visible in Microsoft MAI’s Office prompts and efforts to wrest AI from big labs. If agents become portable across models, the application platform gains power. Vercel wants the second world.

For readers tracking the language vendors use around this shift, our AI Glossary Cuts Through the Jargon Vendors Hide Behind is useful context. The labels matter less than the control point.

Why does Rauch’s price/performance line weaken one-model AI apps?

Rauch’s most direct production argument came down to cost discipline.

“The reality is, when you’re optimizing for production, you start looking at a price/performance, and Gemini models have awesome price/performance characteristics.”

He also said Vercel is seeing “a lot of growth of Gemini,” even though it is “not on the news as much,” and that open models including Deepseek and GLM-5.2 are “taking off.”

That does not mean Rauch is claiming one model wins everywhere. His point is sharper: once AI features serve real workloads, model choice becomes an operating decision, not a brand decision.

Different jobs can justify different models. A coding agent is not the same as an internal sales assistant. A customer-facing agent is not the same as a tool that answers one structured operational question. Rauch’s example of an internal Vercel sales use case is revealing: a sales rep wants to ask which accounts added the most seats in the last two weeks, instead of waiting for a new dashboard project.

The practical implication is that the agent or platform can become the place where model choice happens. Not because abstraction is elegant, but because production teams want cost, control, and flexibility.

The hard part: model swapping is only valuable if the rest of the system still works. Rauch does not give implementation details for that problem in the interview. He names the pieces customers now understand: “model, harness, data platform, sandbox, gateway.” The open question is how reliably those pieces stay interchangeable when the agent is doing real work.

What do Vercel’s token and deployment numbers say about production AI economics?

The source does not provide API price tables, latency benchmarks, failure rates, or margin data. So the cleanest production math here comes from Vercel’s own scale markers:

Rauch also identifies the “two killer apps of agents.” The first is the coding agent. The second is the internal agent that helps run a company.

That second category is where the economics get more interesting. Internal agents need access to company data. They need permissions. They need a record of what they did. They also need limits.

Vercel’s answer is Eve, a framework where agent instructions and skills can be described in natural language, and Vercel Sandbox, which Rauch describes as putting the agent “in a little cage.” The goal is not to stop the agent from acting. It is to constrain what data it can reach and what can leave the sandbox.

This is where Vercel’s position becomes attractive to developers building revenue-generating AI products. If a platform can manage deployment, policy, model access, and agent execution in one place, teams don’t have to rebuild the whole app every time a cheaper or better model becomes usable.

That same workday-friction theme shows up outside AI too. Our look at From Just $13, 5 Desk Gadgets Kill Workday Friction makes a smaller version of the same point: productivity gains often come from removing bottlenecks people tolerate for too long.

Is Vercel trying to beat model labs, or become the layer they all need?

Vercel is not claiming it will build the smartest model. Rauch is arguing for the layer around the model.

That fits Vercel’s business. The company is known for cloud infrastructure that lets developers deploy agents without managing servers. Its incentive is to make AI applications easier to ship, run, restrict, and modify. The model is one input.

The labs have a different incentive. Rauch notes that OpenAI released tools that publish directly to the web “without having to leave the OpenAI enclave.” He calls that “a natural next step” and says it creates “a great opening” for Vercel because users may begin thinking of ChatGPT as a tool for making websites, then ask questions about hosting.

His more important admission is that labs and infrastructure platforms are starting to collide:

“As the models or platforms add more capabilities, they come in direct competition with the infrastructure platforms that already exist.”

That is the control battle. Labs want developers to stay inside their product surface. Vercel wants developers to assemble agents from interchangeable parts.

Rauch’s own ambition is explicit:

“We’re going to be the AWS of this generation, so obviously we’re fighting for a world of open protocols.”

XOOMAR analysis: that line should be read as strategy, not humility. Vercel wants to make model choice boring. If it succeeds, labs still matter, but they compete harder on price/performance because the application layer can switch.

Which side benefits if agents and models split apart?

Different players want different outcomes.

Rauch’s Airbus anecdote shows why enterprises will care. He described a risk where a developer installs the wrong tool and sensitive code leaves for cloud training:

“You have decades of wealth of very specific C++ code for aerospace engineering. Someone comes in and installs the wrong developer tool and boom, all the code goes out to the cloud for training.”

That is why the agent layer can’t just be a convenience layer. It has to become a control layer.

The same logic applies to Rauch’s criticism of SaaS vendors. He says agents are forcing companies to open up because “so many of these SaaS giants build their entire kingdoms on trapping your data, and that’s incompatible with agents.” For a separate example of platform dependency shaping tech strategy, see our coverage of Bookshop.org Kobo Support Revives Its Amazon Fight.

What evidence will show whether model-neutral agents are real or just platform marketing?

The next phase of Vercel AI agents will not be judged by how clean the architecture sounds. It will be judged by whether teams can swap models, keep policies intact, preserve auditability, and still get useful work done.

Evidence that would support Rauch’s thesis:

• Model diversity: More production workloads flowing through Gemini, Deepseek, GLM-5.2, Anthropic, OpenAI, and others based on task fit.
• Enterprise adoption: Internal agents using controlled access to business data without creating unacceptable leakage risk.
• Protocol pressure: Developers demanding open interfaces between models, gateways, sandboxes, and agent frameworks.
• Lab response: Model providers adding more hosting, agent, and publishing tools to keep developers inside their own products.

Evidence that would weaken it is just as clear. If the best agents depend too heavily on one model provider’s behavior, tool system, or hosted environment, portability becomes a sales pitch rather than an engineering reality.

Rauch is betting that production AI will look more like software engineering: modules, libraries, policies, and replaceable parts. The labs are betting, openly or not, that the model can remain the gravitational center. The winner won’t just answer prompts better. It will control the path from user intent to agent action to deployed software.

The Bottom Line

• Vercel is positioning itself as a control layer for production AI, not just cloud infrastructure.
• Separating agents from models could reduce lock-in to any single AI lab.
• The scale of Vercel’s AI usage shows agent governance, security, and model routing are becoming core infrastructure problems.