What happened
OpenRouter announced the next generation of Auto Router on November 6, 2026, and introduced Auto Code (openrouter/auto-code) as a coding-focused option alongside its general-purpose router (openrouter/auto). The company says the new generation was rebuilt on a new technical foundation. In its live test, the routing step was three times faster at the 90th percentile: 0.65 seconds compared with 1.96 seconds. OpenRouter also says the new router outperformed the old one by up to four times, depending on the benchmark. It reports that its existing Auto Router handles more than 2.5 trillion tokens from over 100,000 people each month.
The new routers assess eleven aspects of a request, including task type, difficulty, speed requirements, vision or codebase needs, and whether a larger model, more reasoning, or a cheaper model is appropriate. They combine those judgments with live prices, provider health, cache status, and benchmark performance. The router first narrows the available models according to guardrails, zero-data-retention requirements, provider restrictions, allowed or excluded models, and bring-your-own-key settings. A routing policy then selects a model and reasoning effort; OpenRouter’s separate provider router chooses which provider serves that model. Signals from each turn, including whether the model appears to be struggling and what is already cached, inform the next choice.
Auto Code is tuned for coding agents and harnesses. OpenRouter says it starts with a higher tier of coding models, gives speed greater weight, penalizes failures more heavily, and evaluates measured pass rate, speed, and cost per solved task. Across five benchmarks, the company says Auto Code scored in line with Claude Opus 5.5 while costing 74% less per task. In beta, it stayed with the same model for nearly nine out of ten agent turns, partly to preserve prompt-cache benefits; it can instead adjust reasoning effort when that is sufficient. OpenRouter also reports that, against the previous router, invalid tool calls were four times lower and agentic responses were cut off half as often.
Both routers can be used wherever a model string is accepted. OpenRouter says existing Auto Router users do not need to change anything: openrouter/auto remains the slug, and cost_tier, allowed_models, and excluded_models settings continue to work. There is no additional routing fee; users pay the standard price of the model that handles the request, identified in the response’s model field. The previous router’s drift toward the lowest-cost models led OpenRouter to return Auto’s pricing closer to levels from a few months earlier to include a wider variety of models. Users who prefer the previous cost behavior can set cost_tier to low.
What the archive adds
New Runtime’s earlier coverage described OpenRouter changing its default provider routing to balance cost, latency, throughput, and reliability, while keeping :floor as a price-only option (October 31, 2026 note). That was a provider-selection change; this is a new event at the model-selection layer, with a coding-specific router and reported task-cost results.
Trend signal
This supports the pattern that agent economics is moving toward cost per verified completed task rather than cost per token or model call (Agent economics moves to completed work). Auto Code is explicitly described as weighing pass rate, speed, and cost per solved task. The useful test for teams is whether those reported trade-offs hold on their own coding workloads.
Why now
OpenRouter says experience operating Auto Router for nearly three years exposed weaknesses in using a general-purpose router inside coding harnesses. The new design separates coding traffic and accounts for failures, responsiveness, and cache loss. The company frames the release as a response to that operational feedback and its beta testing.
Connections to other events
New Runtime’s earlier coverage of Baseten Grid described a public-preview control layer combining a model gateway, task-level routing, and guardrails for enterprise coding agents (November 5, 2026 note). Both developments put routing closer to the coding-agent workflow, but OpenRouter’s announcement focuses on automatic model selection and its claimed cost and benchmark trade-offs, rather than an enterprise control plane.
Why it matters
A coding team can try Auto Code by changing the model string rather than building its own model-selection layer. That reduces the setup cost of testing whether different tasks benefit from different model tiers or reasoning effort. The reported 74% lower cost per task is potentially significant, but it is a vendor benchmark claim—not a guarantee for a team’s repositories, prompts, or acceptance criteria.
The router also makes a practical trade-off visible: switching models can discard prompt-cache value, while staying with a poorly matched model can prolong failed work. OpenRouter says it tries changing reasoning effort first and uses session signals to decide when switching is worthwhile. Teams evaluating it should compare completed-task success, human review time, latency, and total spend against their current fixed-model baseline. The benefit is not simply a lower model price; it is whether the router completes the same work reliably at lower total cost.
Limits and uncertainty
The benchmark methodology and task mix behind the five comparisons are not detailed in the announcement, and the 74% figure has not been independently checked here. OpenRouter does not establish that Auto Code will match Claude Opus 5.5 on every coding workload or repository. Its performance and cost will depend on the models available, current prices, provider conditions, cache state, and the team’s routing restrictions. The company says it publishes benchmark results and routing insights, but real-world outcomes still need to be measured against a team’s own tasks.
Skill to build now
Run a small, controlled evaluation before switching a production coding agent. Select representative issues with clear tests or acceptance checks; record the fixed-model baseline for pass rate, completion time, and spend; then repeat with openrouter/auto-code. Keep prompts and harness settings consistent, capture which model was selected, and inspect failed or repeatedly retried tasks. This turns the router’s cost-per-solved-task claim into a decision grounded in your team’s workflow.
Reader takeaway
Auto Code gives coding-agent teams a drop-in way to test automatic model selection, with no routing surcharge. OpenRouter’s cost and benchmark claims are promising, but the decision should come from measured completed-task quality, latency, and total spend on your own work.