The monthly bill had a new line item, and a smaller total. A growing number of companies are adopting software that decides which AI model handles each request, automatically sending simple tasks to cheap models and reserving frontier systems from OpenAI and Anthropic for the work that actually needs them, CNBC reported.
The technology has a plain name: model routing. A routing layer sits between a company’s application and the model providers, scores each incoming request for complexity and dispatches it accordingly. A customer-support query about a password reset goes to a small, inexpensive model; a legal-document analysis goes to the flagship. Procurement teams love the logic, because it is the first AI cost control that does not require anyone to change how they work.
The mechanics are straightforward, and the savings are not trivial. Companies that have deployed routers report that a large share of their requests can be handled by cheaper models, which means the premium-priced tokens sold by frontier labs are being invoked less often. The average revenue per request at OpenAI and Anthropic falls, even as total request volume grows.
CNBC’s analysis drew the conclusion that now hangs over the sector: routing directly erodes the pricing power that underpins frontier AI valuations. OpenAI and Anthropic have built their business models on premium access, with APIs priced for capability and the argument that frontier models pay for themselves. If buyers can separate easy from hard, they will pay for capability only when they need it.
The response, analysts said, will have to come through differentiation rather than uniform pricing. Frontier labs can defend their margins by selling features that routers cannot easily replace: reasoning chains, specialized skills, security guarantees and enterprise support. What they cannot do is charge a premium for simple work, because the market now has a way to measure what simple work is worth.
The economics cut both ways. For the labs, every request routed away from a flagship model is revenue that moves to a cheaper tier, either their own or a competitor’s. For the buyer, the same request is a cost reduction that compounds across millions of calls per month. The two sides are negotiating over a spread that did not exist two years ago, and the routers have made the negotiation visible.
For Anthropic, the timing is uncomfortable. The company has filed confidentially for an initial public offering at a valuation reported at as much as $965 billion, and the pricing question sits at the center of the deal. A strong gross margin will validate the valuation; a margin eroded by routing will raise questions. The IPO is, among other things, a referendum on whether frontier pricing survives contact with corporate cost-cutting.
The routing companies are the quiet winners. Software that manages model selection is a small but fast-growing category, and the larger enterprises that deploy it tend to stay with it once the savings show up in the budget. OpenRouter and Not Diamond are among the better-known names in the space, and each pricing change by the frontier labs tends to drive more interest in the arbitrage.
The trend also changes the shape of demand. Frontier labs have assumed that scale means more premium usage; routing inverts that assumption, pushing premium usage toward the hard problems and commodity usage toward the cheap models. That split has consequences for training decisions, model sizes and pricing structures across the industry, and it rewards labs that can field a wide range of models at different price points.
There is a parallel in the market’s reaction to AI spending generally. Investors have begun scrutinizing the return on AI capital, and tools that cut token bills are popular precisely because they make the returns legible. Every dollar saved by routing is a dollar of revenue lost to the labs, and the asymmetry explains why the labs have been quiet about the trend while enterprise buyers talk about it openly.
The labs have options. They can make their cheapest models so good that routing barely matters, or they can bundle premium features in ways that make the premium stick, or they can price by outcome rather than by token. All three are possible; none is certain. In the meantime, the routers have become one more force shaping the financial futures of the two most valuable private companies in AI.
For the companies watching from the IPO queue, the lesson is that capability alone no longer sets prices. The market for AI tokens is maturing into a market like any other, with buyers shopping, comparing and optimizing. The premium that frontier labs have enjoyed was built on scarcity; routing is the first systematic challenge to that scarcity, and its spread through enterprise accounts is worth watching as closely as any model release.
