← Back to feed

OpenRouter has replaced the benchmark-based logic inside its Auto router with a system that learns from where developers actually spend money. The update to openrouter/auto — live for all users as of August 10 — draws on 55 trillion tokens routed weekly to build a continuously updated picture of which models developers choose for which tasks at which price points.

The previous router used benchmark rankings to make model selection decisions. The new approach inverts the logic: rather than asking which model scores highest on evaluation suites, it asks which model the market has collectively decided is worth paying for in production. OpenRouter calls this the “wisdom of the market.”

How It Works

The router classifies each incoming request by task type, then references the prior seven days of OpenRouter spend data to select the model that market behaviour suggests is the best fit for that task and cost tier. Because the index rolls on a weekly window, newly released models become eligible for routing almost immediately after they attract meaningful traffic, without waiting for evaluation results.

The company tested the approach in beta with thousands of users over several weeks before rolling it out to all accounts. In published benchmark comparisons, the market-spend router outperforms the prior version across tasks including reasoning, coding, instruction following, and data analysis at multiple cost levels.

Early-access users can target openrouter/auto-beta to receive future iterations of the router before general release.

Why It Matters

The approach sidesteps a known problem with benchmark-based routing: leaderboard rankings reflect controlled evaluation conditions, not production performance on real developer queries. A model that tops SWE-bench under a specific harness may not be the best choice for a production agent loop at $0.50 per task.

OpenRouter sees enough token volume — 55 trillion per week — that spend patterns are statistically meaningful at fine granularity. A model that wins majority spend in, say, multi-turn customer support queries at under $1 per task is revealed preference, not benchmark arithmetic.

The shift also makes the router self-updating. When a new frontier model launches and captures developer spend within days, the router adjusts without a manual update. When a model degrades in quality or raises prices and developers move away, the router tracks that too.

What Stays the Same

The model catalog, pricing, and all existing API endpoints remain unchanged. Developers calling openrouter/auto get the new routing logic automatically. There is no migration required.

The full list of eligible models and the spend data that informs routing are not published, but OpenRouter notes that any model with sufficient token volume in the trailing seven days is eligible for selection.