Weave Router

Routing coding-agent requests to the cheapest capable model

Coding agents get a model router in Weave Router, made by Weave Engineering Intelligence, a Y Combinator-backed company with a main product that measures engineering work. Version 2.0 came out on 16 September 2026. The router stands in the path from a coding agent to the model providers. At every turn the agent takes, a classifier judges how hard the task is. The request then goes to the least expensive model able to answer it, and the top-tier models are held back for turns where they matter.

Subscription-aware and cache-aware

The router knows which plans a team already holds. A team can use one vendor's models inside another vendor's agent, routed to whichever plan still has quota left. It also takes the prompt cache into account: a conversation is moved to another model only if the saving outweighs what it would cost to rebuild that cache.

Weave's published comparison

Weave published a run against a frontier coding model on the same Codex harness, with the same 66 tasks, two attempts each, paired on the same day. The router solved 62.1 percent of tasks against 60.6 percent, at 20.5 minutes per trial against 44.1, and at roughly half the cost per trial. Two further benchmarks, Terminal-Bench 4.0 and SWE-Atlas, are mentioned in its launch post. These are Weave's own figures from its own harness, and its product pages promise faster cycles at half the cost with frontier quality.

Who it suits

Weave built it for engineering teams whose coding agents draw on several paid model plans. Weave describes coding harness support and model selection on its product pages. The router has no published price and is sold by talking to the company. Several unrelated products share the Weave name, and this router is a different thing.