OpenRouter has rebuilt its Auto Router from the ground up and launched a brand-new coding-specialized router, Auto Code, the company announced on 6 November. The new generation, built on an entirely new technical foundation after three years of running the original, claims to outperform the old router by up to 4ร— depending on the benchmark, with a routing step about 3ร— faster at p90 โ€” 0.65 seconds against 1.96 seconds in live tests.

The scale behind those numbers is large: Auto Router sees over 2.5 trillion tokens from more than 100,000 people every month, so the new stack has to hold up in the real world, not just on benchmarks. The launch includes the rebuilt general-purpose Auto Router (openrouter/auto) alongside the new Auto Code (openrouter/auto-code). Trying either is a one-line change: set model to openrouter/auto or openrouter/auto-code in any request, or pick them in OpenRouter's chatroom.

The old router made decisions from a couple of imprecise signals, such as task type โ€” a limit imposed by the need to keep latency low and judgments consistent. The new routers weigh eleven factors about each request and combine them with live pricing, provider health and cache status. Governance is built in from the start: guardrails, zero data retention requirements, provider restrictions, allowed model lists and bring-your-own-keys all shape the model pool, and the router picks the cost-optimal model inside it.

Auto Code exists because of the most common complaint about the old router: how it performed in coding harnesses. It is tuned from the ground up for coding work, drawing on a higher tier of coding models and optimising for latency. Across five benchmarks โ€” Deep SWE, SWE-Atlas QA, Terminal Bench 4.0, ฯ„ยณ-bench Banking and MMMU Pro Vision โ€” OpenRouter says Auto Code scores in line with Claude Opus 5.5 while costing about 74% less per task. In beta testing against the previous router, invalid tool calls were 4ร— lower and agentic responses were cut off half as often.

Three design choices stand out. First, the router is restrained about switching models, to protect caching benefits: switching costs extra while the new model rebuilds the input token cache, so it favours changing reasoning effort when that is enough and switches only when the gain is worth the cache it gives up โ€” in beta it stayed on the same model for nearly nine of every ten agent turns. Second, it adapts to failures: when a session is going wrong in ways that point at the model, such as the same error coming back or tests failing after an edit, those signals raise the odds of a switch. Third, it is tuned on real usage as well as benchmarks, verified on production workloads through typesafe/jev-router and openrouter/auto-beta while respecting zero data retention.

Transparency is part of the package too. Pick Auto Router or Auto Code in the chatroom and a routing insights panel appears beside each reply, showing why the router picked that model; the same analysis is available via the API with router metadata enabled. The model pages for both routers now show the live routing mix โ€” which models each router picks most often, split by session type, and how the mix shifts as difficulty rises โ€” and benchmark results are published on OpenRouter's model routing benchmarks page, re-run periodically.

Existing openrouter/auto callers need no slug change, but they should know the behaviour has shifted: the new default returns to pricing closer to where Auto was a few months ago, after the old router drifted toward the cheapest models and left out many exceptional ones. If you preferred the older cheap bias, set cost_tier=low. And routers cost nothing extra โ€” you pay the standard price of whichever model serves the request, and the response's model field tells you which one it was. OpenRouter is candid about the trade-off: no router beats the single best frontier model at every point on the cost and quality curve. Routers earn their place when the work is varied, when costs need to stay steady, and for people who would rather choose once and stay focused on the work itself.

Sources