New AI router learns from execution feedback to cut model costs

A new open-source framework called Agent-as-a-Router promises to make AI model selection smarter and cheaper. Ben Dickson reports for VentureBeat that the concrete implementation, called ACRouter, beats both static routing systems and the common strategy of defaulting to premium models like Claude Opus.

Companies use model routing to send simple tasks to cheap, fast models and complex ones to expensive frontier models. Until now, this routing relied on fixed rules or classifiers trained on past data. Both methods share a weakness: they never learn whether the model they picked actually succeeded at the task.

ACRouter wants to close this gap with a Context-Action-Feedback loop. The system checks its memory for similar past tasks, picks a model, executes the task, and then records whether it succeeded or failed. This feedback shapes future decisions. If an open-source model hallucinates a database column and breaks a SQL query, the router remembers and sends similar future queries to a stronger model.

Cheaper without sacrificing quality

In tests on roughly 10,000 coding and agentic tasks, no single frontier model won every category. Static routers kept sending niche tasks to the wrong models because they had no way to detect failures. ACRouter, by contrast, adjusted routing after each failure signal.

On in-distribution tasks, ACRouter cost $13.21 for the full run, compared to $34.02 when always using Opus, a 2.6 times saving. The approach works best for verifiable tasks like coding, where success or failure is clear. For subjective work like creative writing, the researchers say it offers no benefit. The code and model weights are available under an Apache 2.0 license.