The Information reported on October 15 that Cognition has tightened the amount of model usage it hands to its own engineers, routing expensive calls to a smaller group and pushing everyday work onto cheaper ones. Other large employers are weighing similar rules, according to people familiar with the matter, as the cost of running agents becomes visible in monthly budgets.
Cognition, whose flagship product is the Devin coding agent, has limited high-consumption models to its most efficient engineers. Routine requests are handled by smaller and less costly models, and a routing layer inside the company directs the hardest problems to the expensive ones. The arrangement inverts the pitch that has driven the coding agent business for two years, which was that more AI usage produces more code.
The company is now measuring output per unit of token. That is a different question from whether a model can write software, and it is the one that finance departments have started asking as bills arrive. An agent that generates a large volume of code is not obviously cheaper than an engineer who writes less, if the code has to be reviewed, corrected and maintained by people whose time costs far more than the model calls.
Analysts said the change reflects a shift that is spreading through large technology organizations. Model calls used to be small enough to sit inside a general cloud budget, and they are now large enough that teams track them the way they track storage or compute. Routing requests by difficulty, cost or expected value has moved from an optimization that vendors described to a control that customers actually impose.
The routing layer is the mechanism. A request that resembles thousands of earlier ones can be handled by a small model without a visible drop in quality, while a request that requires reasoning across a large codebase may be worth a premium price. Placing that decision in software rather than in a policy document means the constraint applies continuously and does not depend on an engineer’s judgment about whether a task is hard.
There are limits to how well any router can predict that. Estimating in advance which requests will benefit from a more capable model is difficult, and the errors run in both directions: a task sent to a cheaper model may fail and be retried on a better one, costing more than if it had been routed properly at the start. Retries are where token budgets quietly disappear.
The change also alters how a coding agent company talks about its own product. If usage can be rationed and still deliver the same results, then the value of the agent rests on the quality of the routing and the tooling around it rather than on the scale of the underlying model. That is a more defensible position against competitors who rent the same models from the same providers, and it is a harder one to demonstrate in a demonstration.
Cognition has not said how much it has cut spending, how it defines its most efficient engineers, or what share of requests reach the expensive models. Those figures would show whether the company has found a way to keep the quality of its output while lowering its cost, or has simply decided to accept less capability for a lower bill. The distinction matters to every company that has promised its investors that AI productivity would show up in margins.
Vendors are responding to the same pressure from the other direction. Model providers have introduced cheaper tiers for routine requests and premium tiers for difficult ones, which effectively hands customers the routing decision and the bill that comes with it. Companies that adopt those tiers are making a policy choice about quality, and they are making it in engineering meetings rather than in procurement negotiations.
Internal routing also changes how tools are bought. A team that can send a share of its work to a smaller model can serve more users on the same contract, and a vendor whose product is expensive per request looks worse than one that spreads its calls across several. Sales conversations that once centered on capability now include questions about token consumption per completed task, a measure that most vendors did not publish a year ago.
The result is that a rationing decision made inside one company is likely to be copied. If Cognition can route most requests to cheaper models and report the same output, other firms will ask why their own bills are not structured the same way, and the burden of proof shifts to those arguing for the expensive path.
What is clear is that the cost of running agents has become significant enough to be managed by engineering teams rather than only by finance. Companies that bought AI tools on the assumption that usage would be free are now writing rules about when a model may be called, which users may call it, and what a task must be worth to justify the expensive path. The token budget, once an abstraction in a vendor’s pricing page, has become a line item with an owner.
