Rafay announced on November 5, 2026 that Token Factory customers can set spending policies that switch AI requests to a cheaper model when a budget runs out. The company says its new Token Router and Token Budgets capabilities are available to enterprises and providers operating AI services.
For teams responsible for those services, the question is whether keeping an application running matters more than stopping new work. A configured budget is not necessarily a shutdown instruction.
Three choices at the limit
According to Rafay’s announcement, administrators can assign daily or monthly monetary budgets to an organization, project or individual. Each request is checked against applicable budgets before being accepted, and a user’s allowance follows them across projects. Operators choose the response to exhaustion:
Reject further requests.
Continue serving requests.
Switch to a lower-cost model.
Rafay’s product page describes premium-model access lasting until a team reaches its allowance, followed by service on an economy model. The platform can route between external providers and models run on the operator’s own infrastructure through one connection point. That keeps the application’s destination stable; it does not establish that the models give equivalent answers.
Cheaper service still needs a budget
The practical distinction is between limiting premium-model use and capping all expenditure. A cheaper request still costs money. Once an allowance is exhausted, further paid requests cannot fit within that same allowance. Customers therefore need to establish which budget covers service after the switch before treating it as a hard total-spend ceiling.
Rafay presents continued service and predictable spending as benefits. Its public announcement does not settle how requests already running, simultaneous requests or fallback charges affect a final bill. Refusing new work is the policy to examine when stopping matters more than continuity, but the accounting details still need confirmation.
This is distinct from identifying who incurred the costs, the subject of RohitAI’s report on Kong’s AI cost tracking. Rafay’s announcement addresses what the service does when a threshold is reached.
There is no fixed Token Router feature price in the cited public materials. Rafay’s pricing page lists a platform fee plus GPU/CPU or server-based charges, while its FAQ directs buyers to a tailored quote. Self-hosted fallback still needs infrastructure and operating costs counted alongside any model-provider charges.
Before enabling switching, evaluate the alternative on the application’s actual tasks and confirm it is allowed to receive the same data. The operating decision is not simply which model is cheaper, but whether that model is an acceptable substitute for each workload.