An AI tool that summarizes the occasional contract can go down for an afternoon and nobody files a ticket. Put the same model call inside order processing or a customer service queue, and a price change or a provider outage becomes an operations incident with your name on it.

I see more chief information officers treating model choice as a procurement line item when it behaves like an architecture decision. A contract signed with one model provider shapes your cost curve and your recovery plan for years. You have a cheaper hedge this fall than you had in January, and the usage data shows why.

A pilot budget hides what a finished task costs

Most model providers bill by the token, the unit of text a model reads and writes. A pilot with fifty users and short prompts produces a bill that looks like a rounding error. The same workflow at production volume reads long documents, chains several model calls per request, retries the failures, and runs without an employee pressing a button, so usage climbs faster than headcount.

The price per million tokens tells you little about the cost of a completed insurance claim or support case. Ville Lehto, vice president of strategy at Aiven, told The Register his team tracks cost per task for that reason, because models with different per-token prices finish the same job with very different token counts.

Before you scale, give each critical workflow a per-task cost estimate that includes retries and human review, name an owner who acts when spending passes it, and test what the system does when the budget runs out. It can stop, queue the work, route to a cheaper model, or hand the task to an employee. Pick one on purpose.

Open-weight models took the majority of token volume among OpenRouter's top 50 models

Open-weight models publish their trained parameters. Subject to the model's license and hardware requirements, you can run them yourself or use a hosting provider that serves them. OpenRouter, which routes developer traffic across hundreds of models, reports that open-weight models handled 72% of tokens among its top 50 models in late September 2026, against 27% in October 2025 (OpenRouter, 2026). Closed-model volume grew almost tenfold over the same period, so both camps grew and open-weight models grew faster. These figures describe OpenRouter's classified top-50 traffic, not the wider AI market, and exclude unclassified stealth models.

Price drives most of it. OpenRouter puts the token-weighted list price of closed models at about four times that of open-weight models, and attributes the gap to dozens of providers competing to serve the same open model. The same post reports that the best closed models still score about 11 points higher on the Artificial Analysis Intelligence Index, a composite benchmark of reasoning and knowledge tasks (OpenRouter, 2026).

For an enterprise, the useful part of that data is the second option it creates. A workflow that can call an open-weight model hosted by several providers has somewhere to go when its primary provider raises prices or goes down.

Aiven plans with frontier models and executes with open ones

Aiven, the Finland-based managed database company, gives a working pattern. Lehto told The Register the company faced a choice between strict token budgets for its heaviest internal users and a cheaper way to serve them. It built an internal gateway on the open source LiteLLM project, added budgeting and access controls, and runs open-weight models on rented graphics processing units and its own hosting. Aiven reports that spending among those heavy users fell about 80%, by its own measure (The Register, 2026).

Aiven kept frontier models for the step where quality matters most, planning the work, and sends execution to open-weight models. OpenRouter recommends a similar default from its own data: start on an open-weight model, test it on your tasks, and pay for a closed model only on the steps where the test shows a better result. OpenRouter also describes the reverse arrangement for resilience, with a closed model as the primary and an open-weight model as the fallback that keeps an application answering during a vendor outage.

Both patterns need routing logic between your applications and the models, and a gateway your team controls is one way to centralize it.

NVIDIA is betting the next agent problem is containment, and your team still writes the rules

NVIDIA CEO Jensen Huang spoke at Fireworks AI's Forge event in San Francisco on November 3 and argued that most business value sits in specialized models, built by adapting an open frontier model with your own domain knowledge and evaluation. He did not dismiss the alternative. Huang told the audience, "There's a solid place in the world for closed frontier models," according to The Register.

Huang also predicted that OpenShell, NVIDIA's sandbox runtime for agents, would be the agent story people discuss next year. NemoClaw, NVIDIA's open source reference stack, runs agents inside OpenShell sandboxes and can route them to a locally hosted model, with policies that govern which files the agent can read and which network destinations it can reach. NVIDIA says OpenShell's privacy router keeps sensitive context on local open models and sends requests to frontier models only when the operator's policy allows.

A self-hosted agent in a sandbox lets your team inspect and configure the runtime boundary. With a hosted agent, your visibility and control depend on the provider's features and contract. You inherit the rest of the job, from hardware and patching to writing the policy file itself. Self-hosting can reduce dependence on a provider's per-token pricing while making your team responsible for infrastructure costs and operations, so I would reserve it for workflows where data sensitivity or heavy dependence on one provider justifies the staff time.

Your content platform decides how easily you can change models

The model is the replaceable part of an AI workflow. The documents, permissions, and metadata it reads are not, and the layer that holds them determines how much rework a model change requires. In its annual report, Box describes Box AI as model-neutral, powered by models from several vendors including Microsoft Azure OpenAI, Google Cloud Vertex AI, Anthropic, and Amazon Bedrock, and says it plans to let customers bring their own model. The same filing states that Box does not let partners train their models on customer data (Box, 2026).

That design puts the swap point inside the content platform, where permissions already live, instead of inside every application your developers build. I attended BoxWorks 2026 in San Francisco at Box's invitation, and the questions I would bring to Box or any content vendor are practical ones. Ask which models your configuration uses today, what changes when you switch, and whether the bring-your-own-model option will accept an open-weight model you host.

Generated code stays put, and runtime calls keep the dependency alive

Separate two kinds of exposure when you inventory your AI use. Code a model wrote, that your developers reviewed and committed, keeps running if the model that wrote it disappears. An application that calls a model every time a customer submits a form depends on that provider for as long as it runs.

Sending data to an external model does not by itself transfer ownership of your intellectual property, and your contract and the provider's retention terms set the real exposure. The harder lock-in is architectural: prompts and evaluation suites tuned to one model's behavior. If switching providers means rebuilding the application and retesting every result, your negotiating position at renewal is weak no matter what the contract says. Where no tested alternative model exists for a runtime call, a documented manual process is the minimum fallback.