Weave Router 2.0

What is Weave Router 2.0?

Weave Router 2.0 is a model router for coding agents that plugs into Claude Code, Codex, and Cursor as a drop-in endpoint, reads every agent turn, scores its complexity, and routes it to the cheapest model that can get it right. It can route across existing subscriptions (Claude models inside Codex, GPT models inside Claude Code), is cache-aware so it only switches models when savings beat prompt-cache rebuild cost, and escalates stalled or looping tasks back to frontier models. It is built by Weave Engineering Intelligence (YC W25), used by startups and Fortune 100 companies per the company's Product Hunt launch.

Key features

• Per-turn complexity classifier, trained on 10x more real agentic coding sessions than version 1.0, scores each request in single-digit milliseconds
• Cache-aware model switching: routes only when expected savings exceed the cost of rebuilding the provider prompt cache
• Escalation classifier detects stalls, loops, or misses and hands the task back to a frontier model
• Multi-subscription routing: Claude models inside Codex and GPT models inside Claude Code, draining flat-rate quota first
• One-command setup (npx @workweave/router) that detects and configures Claude Code, Codex, and Cursor
• Vendor benchmark (pass@2): Terminal-Bench 4.0 62.1% vs GPT-6 Astra 60.6% at $5.22 vs $10.03 per trial, 2.2x faster; SWE-Atlas 62.1% vs 66.1% at $2.31 vs $5.04 per trial, 2.5x faster
• Hosted and self-hosted options: source available under Elastic License 2.0, public repo at github.com/weave-os/router with reproducible benchmark harness
• Per-turn visibility of routing decisions and quota savings in the UI; custom cost-vs-speed preferences available besides the auto default

Who it's for

• Teams running coding agents daily who want lower model bills without changing their tools
• Organizations holding both Anthropic and OpenAI subscriptions that want to use whichever plan has quota left
• Engineering leads who want agent traffic routed by task complexity instead of one frontier model for everything
• Enterprise teams (50+ engineers) that want forward-deployed-engineer setup help on their agent traffic

When not to use it

It targets coding agents specifically, not general LLM API traffic from your own apps — a general gateway like OpenRouter fits that better. Early Product Hunt reviewers also noted limited post-hoc visibility into why a specific turn was routed to a weaker model, a gap if every routing decision must be auditable.

FAQ