OpenRouter has rebuilt Auto Router, the feature that decides which AI model answers each request a developer sends, and added a second version called Auto Code for programming work. The company announced both on its blog on 6 November 2026 and says the new Auto Router outperforms the old one by up to 4x.
A router here is a dispatcher. OpenRouter is a marketplace that gives developers one connection to more than 600 models from many labs. Instead of naming a model, you type openrouter/auto where the model name would go, and software weighs how hard the task looks, whether speed matters, and what each model costs right now. Then it picks for you. A separate system, which OpenRouter calls provider routing, chooses which company’s servers actually run the chosen model.
The “up to 4x” needs care. According to OpenRouter’s blog post, the new router beats the old one by that margin “depending on the benchmark”, which makes it the best result across its tests and not an average. The text does not say which benchmark produced it or what was being measured. The company sells the product and ran the tests, and the post cites no independent evaluation.
The other figures are easier to pin down, though they also come from OpenRouter. In a live test, the step where the router decides took 0.65 seconds at the 90th percentile, meaning nine in ten requests were faster, against 1.96 seconds for the old router. During a beta period, invalid tool calls ran at a quarter of the old rate, and answers got truncated half as frequently. That second 4x is a different measurement from the first one.
Auto Code is the more specific bet. OpenRouter says the most common complaint about the old router was how it behaved inside coding tools, so the new version leans toward a somewhat pricier tier of models, puts more weight on speed, and switches away faster when a session goes wrong, for example when the same error keeps coming back or tests fail after an edit. Across five benchmarks, the company reports that Auto Code matches Claude Opus 5.5 on results while costing 74 percent less for each task.
Switching models mid-session has a price of its own. The new model has to rebuild the cache of text it was already given, and that costs extra. So the router changes its reasoning effort when that is enough, and in the beta it kept the same model for close to 90 percent of an agent’s turns, by the company’s count.
One change will show up on invoices. OpenRouter says the old router had drifted toward the cheapest models, which left many strong ones unused, so this release returns to pricing closer to where it was a few months ago. Developers who liked the cheaper behavior can set the cost_tier option to low. OpenRouter says the bill is whatever the chosen model normally charges, and each response names the model that handled it.
The company is blunt about the limits. OpenRouter concedes that no router, its own included, beats the strongest single model across the whole trade-off between cost and quality. Its pitch is for varied workloads, steady spending, and people who do not want to follow every model release. OpenRouter says it handles more than 2.5 trillion tokens (the chunks of text models read and write) from over 100,000 people each month, which gives it plenty of real traffic to learn from.
Anyone already pointing an agent at openrouter/auto was moved to the new router without changing a line, so last month’s spend is the baseline to compare against this month’s.
Reported by the OpenRouter blog on 6 November 2026.
