Model Routing Powered by Wisdom of the Market
By Steven Van ·
Our new Auto router is informed by the model decisions of millions of people, and it outperforms conventional task-based classifiers across a wide…
OpenRouter has rolled out a new version of its Auto router, which picks a model for each request based on what the rest of the OpenRouter community is actually spending on for that type of task, rather than a fixed classifier. It uses the past seven days of usage data across OpenRouter's more than 55T in weekly token spend, so it updates automatically as new models launch and usage shifts.
- Send "model": "openrouter/auto" as before; a lightweight classifier sorts each prompt into one of roughly 30 task types, then ranks models by real-world spend share for that task.
- A new cost_tier parameter (low, medium, high, xhigh, or max) sets how much to spend: low sticks to the cheapest capable models, max draws from the priciest, most capable ones.
- Multi-turn conversations stay on the same model until it's no longer a leading choice for the task, to avoid unnecessary cache rebuilds.
- The old cost_quality_tradeoff parameter still works for backwards compatibility and takes precedence if both are sent.
In OpenRouter's own benchmarks across knowledge, agent, search, research and coding tasks, the new router matched or beat the old one at the default tier while generally costing less, and reached higher scores at the max tier. Results are published on the announcement page, which OpenRouter notes will shift over time as community usage changes.
