Why every company wants an AI model router right now

No one appreciates unexpectedly high bills, yet many companies have encountered such surprises this year while using AI coding agents like Claude Code and Codex for lengthy, autonomous tasks. These agents can operate for extended periods, repeatedly contacting powerful models and accumulating millions of tokens, leading to substantial inference costs.

A recent study reveals that 62% of organizations experienced such unexpected AI expenses in the past year, with 40% requiring board-level attention, 33% implementing emergency spending freezes, and 25% delaying or canceling AI initiatives. This surge in AI model routers, software that enables companies to select the most cost-effective model for specific tasks, has become a hot area in enterprise technology.

These routers can cut inference costs by double-digit percentages, sometimes up to 30%. Startups like OpenRouter, which reportedly attracted acquisition interest from Stripe at a valuation of up to $10 billion, and Not Diamond, which automatically routes requests to the most suitable model for each task, have seized this opportunity.

Other companies, including LiteLLM and Salesforce, are also developing routing capabilities. While cost control is a primary driver for AI model routers, companies are increasingly recognizing their potential to address other concerns as well. Salesforce's president and chief architect, David Ward, envisions routing decisions expanding to incorporate trust, compliance, governance, and measurable business outcomes.

This shift could enable companies to harness the flexibility to switch AI providers or leverage more capable open-weight models if access to certain models changes.