OpenAI and Anthropic face AI overspending, so companies deploy model routing to cut costs

For years, the default strategy in AI has been simple: if you need intelligence, use the strongest model you can get. That approach is getting pressure from a new operational reality. CNBC reports that companies are shifting away from “run everything on the most powerful model” toward a practice called model routing, where each task is matched to the right AI model.

Why does this matter right now? Because the move is framed as a fix for AI overspending, and CNBC specifically calls out the implications as a problem for OpenAI and Anthropic. If customers can route routine or less sensitive requests to cheaper models, the overall pattern of demand can change quickly. In plain English, the buyer behavior becomes more granular, and the “always buy the biggest model” assumption weakens.

Model routing is, at its core, a cost-control mechanism. Not every part of an application needs the top-end model. Some tasks require reasoning depth; others need quick classification, extraction, summarization, or templated generation. When companies route each task to the right model, they can save money without necessarily giving up quality on every single step. That is the key appeal described in the CNBC piece: it addresses AI overspending by reducing unnecessary usage of the most expensive systems.

This is also a governance and operating model story, not just a technical one. Many AI deployments quickly become a “metered budget” problem. Usage grows with adoption, but costs can grow faster if teams route everything to the highest-cost option. Model routing changes that operating dynamic by making cost a controllable parameter. For decision-makers, the question becomes how to measure “right model” selection well enough that savings are real and reliability does not degrade.

There is a second-order implication that boards and senior finance leaders tend to care about: vendor concentration risk can shift. If a company no longer treats the top model as the default for every workflow, then procurement patterns can become more diversified even when the primary vendor remains important. The buyer may still call OpenAI or Anthropic for the tasks that truly require frontier capability, but route other tasks elsewhere or to smaller internal models. That can affect forecasting, revenue mix, and how customer “land and expand” plays out in practice.

Another angle is that model routing tends to favor organizations that can engineer and operate their AI stack tightly. Frontline product teams might want to ship fast with the biggest model and move on. Model routing requires more design: deciding what tasks should route where, monitoring performance, and iterating over time. That means operational maturity becomes a competitive advantage. If some companies can implement routing quickly and consistently, they can spend less per user, reach profitability earlier, or underbid competitors on cost-to-serve, depending on their business model.

Regulatory background matters here mostly through the lens of accountability. As governments and regulators push for responsible AI use, customers increasingly need visibility into how systems behave, how outputs are generated, and how risk is managed. Model routing can add layers to the system, which makes governance more important. The more decision points exist (which model handled which task), the more an organization may need internal controls, auditability, and documentation practices. Even if the CNBC piece is focused on overspending, the operational shift can ripple into compliance workflows.

Finally, the strategic stakes for OpenAI and Anthropic are clear in the framing. CNBC positions the routing approach as a problem for them because it challenges the economics of always defaulting to their most powerful models. The competitive question becomes: can they retain demand for the most expensive model where it is truly needed, while also competing effectively for routed “right task, right model” decisions? For other executives watching this, the lesson is that the AI marketplace is not just about model breakthroughs anymore. It is about how efficiently those models are used, how cost curves bend with routing, and how fast buyer strategies can change after an “overspending” diagnosis.

In short, model routing is a practical response to AI overspending, and it is reshaping how customers allocate usage among frontier providers. That is good news for budgets. It is a wake-up call for any lab whose revenue assumptions rely on the idea that customers will route everything to the biggest brain available.