OpenRouter rebuilds Auto Router and adds Auto Code: routing step drops to 0.65s at p90 from 1.96s

OpenRouter launched a next-generation Auto Router and, alongside it, Auto Code, a router tuned from the ground up for coding work; both run on the same routing engine. The previous Auto Router relied on a couple of imprecise signals, such as task type, to drive model selection, a limit required to keep latency low and judgments consistent. Across five benchmarks Auto Code scores in line with Claude Opus 5.5 while costing 74% less per task, and the routing step runs at 0.65s at p90 in a live test against 1.96s for the old router.

The router first narrows the model pool by governance policy, then a fast decision model rates the request, and finally a routing policy chooses the model and reasoning effort. The decision model rates eleven things about each request in one call, such as task type, whether the task just changed, difficulty, whether speed matters, whether vision or codebase knowledge is needed, and whether a bigger model, more thinking or a cheaper model would serve better. The routing policy weighs live prices, cache status and benchmark performance, the request then goes through the same provider routing as a direct request to that model, and each turn's results inform the next turn's model selection.

The demonstrated scope is model routing on the OpenRouter platform: Auto Router targets mixed traffic and general agents, Auto Code targets coding agents and harnesses, and both choose only inside the model pool defined by the user's governance policies, zero data retention requirements and provider restrictions. The releasing organization states that no router on the market beats the single best frontier model at every point on the cost and quality curve, and the new Auto Router is no exception.

The releasing organization notes that it is easier for a router to score well on benchmarks than to perform well in practice, and says Auto Router sees over 2.5T tokens from over 100k people every month. In its beta, against the previous router, invalid tool calls were 4× lower and agentic responses were cut off half as often. This release also returns pricing closer to where Auto was a few months ago to introduce a greater variety of models; users who preferred the previous behavior can set cost_tier to low.