OpenRouter launches Auto Code, a router that picks models for coding agents by pass rate, time and cost per solved task

OpenRouter has introduced Auto Code (openrouter/auto-code), a router built specifically for coding agents and harnesses. It uses the same routing engine as OpenRouter's Auto Router, but treats every request as part of an agent session rather than as an isolated prompt.

How Auto Code chooses models

According to OpenRouter, model choices are guided by three measures:

• Coding-task pass rate
• Completion time
• Cost per solved task

The router follows edits, test results, and recurring errors across turns. While work is going well, it keeps the current model. When trouble persists, it tries more reasoning effort before switching to a different model.

The default coding tier starts from stronger models and weighs failed code changes more heavily. Your own model and provider restrictions still apply to routing decisions.

Availability and pricing

Auto Code is not currently available through EU in-region routing. OpenRouter points to its in-region routing docs for details. Pricing for Auto Code is not yet disclosed on the model page, and no context window is listed for it.

The page also shows live usage panels for the last 24 hours (1h and 24h views) and a "current routing mix" view filterable by all tasks, code changes, coding questions, and general. OpenRouter describes the mix as showing how the router would choose for new work.

Related OpenRouter routers

The Auto Code page lists several other OpenRouter routing products:

• Auto Router (experimental version): OpenRouter describes this as the beta version where it tests new improvements before graduating them to the main Auto Router. Responses are priced at the same rate as the routed model. The page lists a 2M context window.
• Pareto Code Router: Keeps a tiered shortlist of coding models ranked by Artificial Analysis coding percentiles. A min_coding_score between 0 and 1 on the pareto-router plugin sets the minimum strength, and the router defaults to the High tier if the value is omitted. A Nitro variant ranks models in a tier by measured throughput and sends each request to the fastest. The page lists a 2M context window.
• Fusion: Runs a panel of expert models in parallel with web search and fetch, then has a judge model synthesize the results. Requests are priced as the sum of the underlying completions. The page lists a 1M context window, and limits depend on the models selected.

What this means

Auto Code differs from a general-purpose router in what it evaluates. It judges routing across a whole agent session, using edits, test results, and repeated errors as signals, instead of scoring each prompt independently. The stated order of escalation, more reasoning effort first and a model switch second, indicates the router is designed to avoid changing models mid-task unless a task stalls.

For teams running coding harnesses through OpenRouter, the practical points are that existing model and provider restrictions carry over, and that EU in-region routing is not currently supported. Because pricing and context window are not listed on the page, cost and limits for Auto Code cannot yet be compared directly with the Auto Router or Pareto Code Router.