Cost shock drives surge in AI model routers for enterprises

For much of 2024 and 2025, enterprise AI adoption followed a simple pattern: deploy the most powerful frontier model and let it run. OpenRouter co-founder and chief operating officer Chris Clark explained that the C-suite was "pounding the table" for AI adoption, but serious technical deployment only accelerated this year as AI agents—autonomous systems that use tools, call external services, and take action—became viable at scale. Claude Code, Anthropic's AI coding agent released in mid-2025, exemplified this shift, offering capabilities far beyond chatbot interactions.

But the economic model broke immediately. Unlike a brief chatbot conversation, agents can operate for hours, repeatedly querying frontier models and accumulating millions of tokens. A developer might deploy an agent, leave for lunch, and return to a bill in the thousands of dollars. The problem was systematic: companies had no cost guardrails and defaulted to the most expensive models for every task.

A new study quantified the impact. Sixty-two percent of organizations said unexpected AI expenses materially altered a business decision over the past year. Among them, 40% required board-level escalation, 33% implemented emergency spending freezes, and 25% delayed or canceled AI initiatives outright. The result was a sudden rush toward AI model routers—software that allows organizations to select the model offering the best balance of cost, speed, and performance for each task. Companies report that intelligent routing can reduce inference costs by double-digit percentages, reaching up to 30% in some cases.

The router market has fragmented quickly. OpenRouter provides a marketplace and unified gateway to hundreds of models; as of the time of reporting, it was reportedly in acquisition talks with Stripe at a valuation of up to $10 billion. Not Diamond, working with enterprise clients like SAP, automatically routes requests to the optimal model based on task complexity and conversation history. LiteLLM lets enterprises build and manage their own routing infrastructure. Large vendors—Salesforce, Databricks—are integrating routing into their platforms. Cursor, Ramp, Meta, and even Runway have announced routers in recent weeks. Despite the crowded field, Clark cautioned that most entrants underestimate the difficulty. Production-grade routing requires deep partnerships with model providers, constant monitoring of thousands of endpoints, rapid response to outages and specification changes, and ongoing validation of data policies and GPU configurations. "It's not a thin software wrapper," he said. "It's a deep partnership."

Enterprises see routing as solving a second problem beyond cost: vendor lock-in and resilience. When Anthropic recently restricted access to its Fable model, Florian Douetteau, CEO of enterprise AI platform Dataiku, noted that customers around the world rethought their dependence on a single provider. A Fortune 500 CIO described discovering late on a Friday that a critical business process's underlying model might not be available Monday. Routers promise flexibility to switch providers or fall back on increasingly capable open-weight models if access changes. Salesforce president and chief architect David Ward outlined a longer vision: routers would eventually control not just model selection but trust, compliance, governance, and measurable business outcomes—determining what data an agent can access and whether a fine-tuned open-source model is sufficient. "It's not just which model for which job," Ward said, "it's which tool, which skill gives me the measured outcome I want."