If you run a coding agent through an API gateway, you’ve probably had the “which model for this?” argument with yourself more than once. On November 6, OpenRouter shipped its answer: a rebuilt Auto Router and a new router built just for coding, called Auto Code.
You opt in by setting the model to openrouter/auto-code (or openrouter/auto for the general router) in any request. After that, the router decides which model handles each turn.
What’s actually new
According to OpenRouter, the old Auto Router leaned on a couple of imprecise signals like task type. The new one has a fast decision model rate eleven things about each request — what kind of task it is, whether the task just changed, how hard it is, whether speed matters, whether it needs vision or codebase knowledge — and combines that with live pricing, provider health and cache status to pick both a model and a reasoning effort.
Auto Code is the coding-specific version. OpenRouter says it favors a slightly more expensive tier of coding models, weighs speed more heavily, and penalizes failures harder so it course-corrects faster.
The part that matters for agents: it remembers the session
This is what makes it more than a price picker. Per the announcement, the results of each turn inform the next one:
- It tries not to burn your cache. Switching models means rebuilding the input token cache, so the router prefers bumping reasoning effort and only switches when the gain is worth it. OpenRouter says that in beta it stayed on the same model for nearly nine of every ten agent turns.
- It reacts to a session going wrong. The same error coming back, or tests failing after an edit, raise the odds of a switch, so a bad early pick is less likely to stick.
- Your governance still applies. Guardrails, zero data retention, provider restrictions, allowed or excluded model lists and BYOK keys decide which models are even in the pool.
The numbers, and how to read them
OpenRouter’s claims: the new foundation outperforms the old router by up to 4x depending on the benchmark; the routing step is 3x faster at p90 in its live test (0.65s against 1.96s); across five benchmarks, Auto Code scores in line with Claude Opus 5.5 while costing 74% less per task; and in beta, invalid tool calls were 4x lower and agentic responses got cut off half as often versus the previous router.
Those are vendor numbers, not independent tests. To OpenRouter’s credit, the post says plainly that no router on the market beats the single best frontier model at every point on the cost and quality curve, and that this one is no exception. It also notes the new routers lean toward pricing closer to where Auto was a few months ago, with a cost_tier set to low if you want the old cheaper behavior.
What I’d check before pointing an agent at it
- Turn on router metadata. OpenRouter says the reasoning behind each pick is available through the API. Log it next to your test results so you can see why a turn went where it did.
- Run your own task set. Same repo, same tests, same tool permissions, routed vs. your current single model. Count solved tasks, retries and cost per solved task, not cost per token.
- Watch the switches. If your harness depends on a specific model’s tool-calling quirks, a mid-session switch is where it’ll show.
- Keep the escape hatch. OpenRouter’s own advice: if you already know the model that works best for your work, keep using it.
For teams here in Mexico paying for tokens in dollars, “cost per solved task” is the metric that matters to your budget, so a router that optimizes for it is worth an afternoon of testing. Just test it on your code, not on the launch post’s charts.
Docs: OpenRouter Auto Router documentation.
Hero photo: Railroad switch by Artyom Svetlov, via Wikimedia Commons, CC BY-SA 4.0; cropped and resized.
