OpenRouter Auto Code is a new model router for coding agents, launched on 6 November alongside a rebuilt general-purpose Auto Router. The company’s announcement says both select a model and reasoning effort for a request, with Auto Code using a policy that gives more weight to coding performance, speed and recovery from failure.
For developers, the choice is between fixing a harness to one model and letting a routing policy choose from eligible models as work changes. OpenRouter says routing adds no extra fee, but the model serving each request is still charged at its standard price. Its claims about lower cost per task and faster routing therefore need their particular baselines, rather than being read as free inference or guaranteed savings on every codebase.
OpenRouter Auto Code gives coding failures more weight
The Auto Code documentation says it starts from a higher cost tier containing stronger coding models than the general router’s usual pool. It weights speed more heavily and penalises failures more harshly, increasing the chance of changing course when tests fail or the same error recurs.
The routing process first narrows the model pool using requirements such as capabilities, zero data retention policies, provider restrictions, bring-your-own-key settings and allowed or excluded models. It then rates the request and weighs measured pass rate, speed and cost per solved task alongside live prices and cache status. The company says the rebuilt engine considers eleven factors about each turn.
This is separate from provider routing. Model routing chooses the model; provider routing then chooses who serves that model using provider preferences and operational information. The selected model appears in the response, and optional router metadata can expose the decision’s reasoning and candidates.
Auto Code also favours changing reasoning effort before switching models, to preserve the prompt cache. A session identifier lets it treat the turns of an agent session as one conversation. The docs describe allowed and excluded model lists as hard restrictions: if none remains eligible, the request fails rather than choosing outside that pool.
These decisions sit below the visible agent workflow. Cursor Live Apps concerns agents building and hosting applications, while Carta Evergreen’s multiplayer coding concerns teams working with coding agents. Neither story establishes an integration with this router; they show why model selection is only one layer of an agent-based development system.
The speed and cost figures describe different comparisons
OpenRouter says the rebuilt Auto Router outperforms its predecessor by up to four times depending on the benchmark. Separately, it reports a routing-step p90, or 90th-percentile, latency of 0.65 seconds against 1.96 seconds in its live test. That is time spent selecting a route, not the total time to generate code, run tests or finish a development task.
For Auto Code, the company reports scores in line with Claude Opus 5.5 while costing 74% less per task across five benchmarks. That is its comparison with a single-model baseline, not the same comparison as the general router’s improvement over its predecessor. OpenRouter also explicitly says no router beats the best single frontier model at every point on the cost-and-quality curve.
Cost per solved task differs from a cheaper token
The useful unit for a coding workflow is not necessarily the price of one token. A cheaper model can require more attempts, while a stronger model can cost more on an individual turn. A policy that weighs failure signals and cost per solved task is trying to balance those possibilities across a session. The headline benchmark saving describes one measured comparison; it cannot settle that balance for every repository, restriction set or task sequence.
The eligible pool is equally important. A team that restricts providers or models is asking the router to optimise within a smaller set, not authorising it to relax those boundaries to chase a score. Cache preservation adds another trade-off: switching may improve the model choice while giving up previously cached input. These design choices explain why the router considers both session history and current conditions instead of simply choosing the cheapest listed model.
Evaluation therefore needs to keep quality, failures, routing delay and the final bill distinct. A faster selection step can help responsiveness without guaranteeing faster completion, and zero additional routing fee leaves underlying usage costs intact. The relevant comparison is the workflow’s result under its actual constraints, not an isolated advertised percentage.
TSN’s Devin Review 2.0 coverage similarly separates a vendor’s code-review cost claims from independent proof of code quality. For this release, OpenRouter’s periodically refreshed routing benchmarks are the place to watch how its published trade-offs change as models change.
Sources
- OpenRouter: Introducing the New Auto and Auto Code Routers, 6 November 2026
- OpenRouter docs: Auto Code
Related stories
Carta Evergreen Opens Multiplayer AI Coding for Teams provides the team-workflow context around model costs, a different layer from choosing a model for each agent turn.
Follow @tsnmediahub on X for breaking AI news as it happens.
Follow on X
