Auto Code

OpenRouter launched its next-generation Auto Router together with Auto Code (openrouter/auto-code), a router specialized for coding agents and harnesses. The rebuilt routing foundation reportedly outperforms the previous router by up to 4x depending on benchmark, with a routing step measured at 0.65s p90 versus 1.96s in live testing, and the router already handles over 2.5T tokens monthly from more than 100k users. Instead of the old coarse signals such as task type, the new generation weighs eleven per-request factors and combines them with live pricing, provider health, and cache status, so it can work with any allowed model set while honoring guardrails, zero data retention, provider restrictions, allowed model lists, and BYOK keys. Auto Code is tuned for coding performance with a higher tier of coding models and latency optimization, and routing decisions can be inspected per model page and in the chatroom.

OpenRouter's new openrouter/auto-code router, announced November 6, picks a model and reasoning effort on each turn for coding agents. Routing weighs pass rate, speed, cost, live pricing, and cache state, and the rebuilt openrouter/auto variant covers general work. Users pay the price of whichever model is selected rather than a router-specific rate. The report attributes to OpenRouter the claim that Auto Code scores in line with Claude Opus 5.5 while cutting cost per task by about 74%. The router is set by assigning the model to openrouter/auto-code in a request, with the announcement read directly on openrouter.ai.

Coverage of the November 6, 2026 launch reports OpenRouter's vendor-stated claim that the coding-focused Auto Code router scores in line with Claude Opus 5.5 across five benchmarks at roughly 74% lower cost per task, with routing p90 latency of about 0.65s versus 1.96s for the prior router. Users set the model to openrouter/auto or openrouter/auto-code and pay the served model's list price, with no additional routing surcharge. The write-up adds that routing weighs eleven request factors plus live price, provider health, and cache status, and that in beta the router stayed on the same model for nearly nine of ten agent turns to protect caches. Compatibility is retained for cost tier, allowed models, and excluded models parameters, and routing reasons are exposed through both the chatroom and the API.

OpenRouter launched its next-generation model routing on November 6, 2026, rebuilding the general-purpose Auto Router and adding Auto Code (model key openrouter/auto-code), a router tuned specifically for coding agents and harnesses. Routing decisions are made per request by evaluating eleven factors in a single call, combining live pricing, provider health, benchmarks, and cache status to choose both the model and its inference effort level. Auto Code shares the same routing engine as Auto Router but is tuned from scratch for coding performance, selecting models by pass rate, completion time, and cost per resolution while tracking edits, test results, and recurring errors across turns. It keeps the current model while work is in progress and raises inference effort before switching when stuck, while reporting p90 routing latency of 0.65 seconds, down from 1.96 seconds previously.

OpenRouter's official model page describes Auto Code (openrouter/auto-code) as a router built for coding agents and harnesses that shares the Auto Router engine but treats every request as part of an agent session. Model choices are guided by coding-task pass rate, completion time, and cost per solved task, and the router follows edits, test results, and recurring errors across turns, keeping the same model while work is going well. The page notes that when trouble persists, Auto Code tries more reasoning effort before switching models, that its default coding tier starts from stronger models, and that user model and provider restrictions still apply. A live routing panel shows the current request mix, including GPT-6.1 Sol at 65.9% of requests, alongside a drop-in OpenAI-compatible API snippet for calling the slug.

OpenRouter's November 6, 2026 release note announces a rebuilt Auto Router alongside a new coding-focused router, Auto Code (openrouter/auto-code), usable by setting the model slug to openrouter/auto or openrouter/auto-code in any request. The rebuilt router is said to outperform the previous version by up to 4× depending on the benchmark, with a routing step that is 3× faster at p90 in a live test, measuring 0.65s against 1.96s for the old router. The new generation replaces imprecise signals such as task type with eleven request factors combined with live pricing, provider health, and cache status, letting the router work with any model set. Governance features include guardrails, zero data retention requirements, provider restrictions, allowed model lists, and BYOK keys, and Auto Code was tuned specifically for use inside coding harnesses.

OpenRouter's model discovery page describes Auto Code as the router built specifically for coding agents and harnesses, using the same routing engine as the general-purpose Auto Router. Each request is treated as part of an agent session, and model choices are guided by coding-task pass rate, completion time, and cost per solved task, with the default coding tier starting from stronger models and weighting failed code changes more heavily. Across turns, the router follows edits, test results, and recurring errors, keeping the same model while work is going well and attempting more reasoning effort before switching models when a task stalls. Model and provider restrictions still apply to Auto Code, so existing governance settings continue to constrain which underlying models it may select for a coding session.