OpenRouter's new Auto Router picks models by real-world spend
The rebuilt router reads a week of community spend to route ~30 task types; OpenRouter also ships web-search benchmarks as Anthropic locks Sonnet 5's price.
Copy markdown
Routing by the wisdom of the market
The rebuilt openrouter/auto classifies each prompt into ~30 task types, then routes by the last seven days of community token spend — over 55T weekly — instead of a static classifier. You choose a cost tier (low, medium, high, xhigh, max) and it picks a primary model plus fallbacks, keeping a conversation on its model until a clearly better one appears.
Cheaper on most jobs, far smarter on research
At default settings the new router hits 62.9% on DSQA research (up from 43.2%) and 61.6% on WideSearch (up from 53.1%), while cutting an MMLU Pro run from $393 to $141 — about 64% less. The tradeoff is small: MMLU Pro knowledge accuracy slips to 85.2% from 86.6%. Reach for openrouter/auto-beta to get the newest tuning first.
Also new: live web-search benchmarks for agents
OpenRouter published benchmarks on Aug 12 comparing search engines, depths, and models so you can pick the right browsing stack for an agent instead of guessing. It pairs with OpenRouter's web search-and-fetch plugin that works across every model.
Elsewhere: Anthropic locks Sonnet 5 at $2/$10
Anthropic's platform release notes confirm Claude Sonnet 5's introductory pricing — $2 in, $10 out per million tokens — is now permanent, and the September 1 rise to $3/$15 will not happen. Many price trackers still list the old hike, so check before you budget for a bump that's gone.