Weave Router 2.0 - Subscription aware coding agent router
by•
Use Claude models in Codex and GPT models in Claude Code on your existing plans, routed to whichever has quota left. Weave Router 2.0 routes each coding agent request to the cheapest model that can get it right. On Terminal-Bench 4.0 and SWE-Atlas it matches GPT-6 Astra at half the cost, 2x faster. Powered by a new classifier that scores task complexity and cache-aware switching that only moves when savings beat the rebuild cost.
Replies
Hey Product Hunt, I'm Adam, Co-founder of Weave!
The Problem
Coding agents send every request to one model. A one-line config change and a cross-service migration both go to the same frontier model at the same price. Teams are paying frontier rates for routine work, and the bill grows with every agent they add.
Most tools in this space follow one of two flawed approaches:
❌ Gateways – They give you one endpoint and a menu of models. You still pick the model yourself, so nothing actually gets cheaper.
❌ Manual model switching – Flipping /model mid-session breaks your prompt cache, and on long sessions the re-billed context eats the savings.
After running into this on our own agent traffic, we built Weave Router to solve it.
How Weave Router 2.0 is Different 🚀
Weave Router plugs into Claude Code, Codex, and Cursor as a drop-in endpoint. It reads every request your agent makes, scores how hard it is, and sends it to the cheapest model that will get it right.
The benchmarks
We ran 2.0 head to head against GPT-6 Astra, each task run twice (pass@2).
📊 Terminal-Bench 4.0 – Tied pass rate within the confidence interval. 52% of Astra's cost. 2.2x faster.
📊 SWE-Atlas Codebase QnA – Tied pass rate within the confidence interval. 46% of Astra's cost. 2.5x faster.
Full results at weaveos.com/router.
🔹 Complexity-scored routing – A new classifier trained on 10x more real agentic coding sessions routes on task difficulty, not task type. DeepSeek, GLM, and Kimi handle the simple turns. Astra and Fable handle the hard ones.
🔹 Cache-aware switching – 2.0 tracks cache state per provider and per session. It only switches models when the expected savings beat the cost of rebuilding the cache. This is where most of the cost win comes from.
🔹 Multi-subscription routing – Use Claude models inside Codex and GPT models inside Claude Code. Route by complexity, by cost, or by whichever plan still has quota this month.
Who is this for?
If you or your team run coding agents daily, Weave Router gets you Astra-level results at about half the cost and more than twice the speed, without changing your tools.
🔗 Get started today
Try it at weaveos.com/router. Install with one command: npx @workweave/router. Source-available under Elastic License 2.0. Teams of 50+ engineers can book time with our FDEs.
@workweave @adambcohen nice launch ! the routing base on the task difficult instead of just the task type of makes a lot of the sense.
this is a really interesting approach.
Hello everyone! 🎉 I'm thrilled that we've launched Weave Router 2.0! 🚀 Many of you have seen us preparing for the launch, and we've done it! Today is a big day for us, as this launch is actually our first public product release.
We would greatly appreciate your feedback, comments, and reviews. 💬
Also, as a gift to all Product Hunt community members, we are offering a 20% discount on routing through the end of 2026!
Dumper cheaper models are known to produce longer trajectories when solving the same task. What are your thoughts on it in the context of smart routing and cost optimisation?
@dmytriis Totally agree! We've run a lot of rollouts on production agent traces to measure how efficient each model actually is. On many tasks (usually complex ones, though not only those) dumber models spin their wheels and burn a ton of tokens, so we try to limit the scenarios where that happens. Artificial Analysis publishes a pretty great tokens-per-task score for every model. We try to go a bit deeper by delineating between task complexity and type as well.
Congrats on the launch. I'm interested to know that when a cheaper model gets a turn slightly wrong, does the router notice and escalate on the next turn, or does the agent carry the mistake forward?
@joy_gu yes it does! it's aware that if it's running and not working as effectively as possible it will escalate it to a more intelligent model.
Very awesome guys - especially now that the 20x plan in codex is temporarily disabled 😵😵😵
@ilyatkachov ahaha 100%
The cheapest model that can get it right approach is really interesting How does Weave decide which model is best for each coding task?
@shivam_kushwaha16 We trained a model on an extremely large dataset that we gathered from customer agent traces as well as model rollouts and continuously improve via RLHF!
Congrats on the launch, guys!
@andrew_roth2 thanks!
Congrats on the launch! This looks really useful. My Question: Is there a free tier to try it out first?