Weave Router is a model orchestration proxy for agentic coding workflows, sitting between tools like Claude Code, Codex, Cursor, and opencode and the underlying LLM providers, then sending each request to the cheapest model capable of producing an equivalent answer. The tool installs with a single command, runs locally, and chooses the best large language model for each request without exposing user data. Version 2.0 pushes the routing quality claim further, with the team asserting better results at lower latency and cost than GPT-6 Astra on Terminal-Bench and SWE-Atlas, an important signal for engineering leaders trying to hold down inference bills as agent usage scales.
Under the hood, this release is driven by an upgraded session analysis model that classifies what an agent is actually trying to do on a given turn, a roughly tenfold increase in training data, and reworked cache eviction logic. The pitch to buyers is straightforward, offering frontier-tier answer quality at the speed and price of much smaller models by matching each action to the right tier rather than defaulting every call to the most expensive option. The router is agent-aware and built for workflows that mix planning, codebase exploration, implementation, and review requests , which is where blanket use of top-end models tends to waste the most spend.
Weave is a Y Combinator W25 company based in San Francisco, founded by Adam Cohen and Andrew Churchill, and recently raised a $13.5M Series A led by Standard Capital . The broader product measures both human and AI engineering output across the SDLC, and the Router is the piece that turns that telemetry into per-request routing decisions for teams already spending heavily on coding agents.
Comments (15)
Sign in to join the discussion.
P
Priya VatsanSep 10, 2026
Beating GPT-6 Astra on Terminal-Bench is a bold flex to put in the first line. Where's the eval methodology link, I want to nitpick before I believe.
M
Morten HauglandSep 10, 2026
The launch video pacing is genuinely tight, whoever cut it earned their coffee. Voiceover a little too NPR calm for a router claiming to dunk on GPT-6 though.
I
intern sznSep 10, 2026
sorry naive question but what happens when the router picks the wrong model and the agent loops forever, do i get charged for that or
K
Kwame OtienoSep 10, 2026
Routers claiming to be faster AND cheaper AND better is the trilemma equivalent of a menu that says gluten free pizza that tastes like real pizza. I'll believe it after a week in prod.
L
Lin Z.Sep 10, 2026
Shipped a model router at Meta in 2019 for ranking infra, we called it Loom. Different problem space obviously but the routing heuristics rhyme.
S
Srivaths R.Sep 10, 2026
Does this hit SOC2 Type 2, and can we route between our private Bedrock deployment and public endpoints under one policy? Procurement wants to know before I can even demo internally.
G
gm.ethSep 10, 2026
Weave Router 2.0 but no $WEAVE token to align incentives between callers and providers? missed opportunity to bootstrap a real routing marketplace.
D
Deniz AksoySep 10, 2026
GPT-6 Astra isn't out yet so beating it is technically the easiest benchmark of all time. Congrats on the undefeated season.
T
Tomasz K.Sep 10, 2026
Been building something adjacent for 3 weeks, was planning to launch next month. Weave shipping this now is either a gift or a gravestone, TBD.
J
Juno OkaforSep 10, 2026
Tagline rewrite, no charge: 'Weave Router. The cheapest path to the right model, every call.' You're welcome.
M
Maeva QuinnSep 10, 2026
Onboarding better be a one-line SDK swap or nobody's ripping out their existing router. First tap matters more than benchmark charts.
Y
yosef b.Sep 10, 2026
The best router is the one you forget you're using. Second best is the one that emails you when it saved you money.
S
Stan VukovicSep 10, 2026
Every router demo looks amazing until you actually plug it into a real codebase with 40 tools and weird retries. Post the trace logs or it's vibes.
A
Amara CalderSep 10, 2026
The tweet copy is doing a lot with two sentences and one chart, respect. Would've killed for a 10 second before/after cost demo instead of just the benchmark bar.
S
Sanju LahiriSep 10, 2026
Portfolio co Fathomline uses a similar routing layer and swears the real moat is the eval harness, not the routing itself. Curious where Weave's leverage actually sits.
Work With Us
Book an intro call to explore our suite of services. We might have a waitlist.
Weave
Engineering and token intelligence to optimize the SDLC
Engineering and token intelligence to optimize the SDLC
Tags
TLVC Rating
Community Rating
No ratings yet
to rate this launch.
Weave Router is a model orchestration proxy for agentic coding workflows, sitting between tools like Claude Code, Codex, Cursor, and opencode and the underlying LLM providers, then sending each request to the cheapest model capable of producing an equivalent answer. The tool installs with a single command, runs locally, and chooses the best large language model for each request without exposing user data. Version 2.0 pushes the routing quality claim further, with the team asserting better results at lower latency and cost than GPT-6 Astra on Terminal-Bench and SWE-Atlas, an important signal for engineering leaders trying to hold down inference bills as agent usage scales. Under the hood, this release is driven by an upgraded session analysis model that classifies what an agent is actually trying to do on a given turn, a roughly tenfold increase in training data, and reworked cache eviction logic. The pitch to buyers is straightforward, offering frontier-tier answer quality at the speed and price of much smaller models by matching each action to the right tier rather than defaulting every call to the most expensive option. The router is agent-aware and built for workflows that mix planning, codebase exploration, implementation, and review requests , which is where blanket use of top-end models tends to waste the most spend. Weave is a Y Combinator W25 company based in San Francisco, founded by Adam Cohen and Andrew Churchill, and recently raised a $13.5M Series A led by Standard Capital . The broader product measures both human and AI engineering output across the SDLC, and the Router is the piece that turns that telemetry into per-request routing decisions for teams already spending heavily on coding agents.
Comments (15)
Beating GPT-6 Astra on Terminal-Bench is a bold flex to put in the first line. Where's the eval methodology link, I want to nitpick before I believe.
The launch video pacing is genuinely tight, whoever cut it earned their coffee. Voiceover a little too NPR calm for a router claiming to dunk on GPT-6 though.
sorry naive question but what happens when the router picks the wrong model and the agent loops forever, do i get charged for that or
Routers claiming to be faster AND cheaper AND better is the trilemma equivalent of a menu that says gluten free pizza that tastes like real pizza. I'll believe it after a week in prod.
Shipped a model router at Meta in 2019 for ranking infra, we called it Loom. Different problem space obviously but the routing heuristics rhyme.
Does this hit SOC2 Type 2, and can we route between our private Bedrock deployment and public endpoints under one policy? Procurement wants to know before I can even demo internally.
Weave Router 2.0 but no $WEAVE token to align incentives between callers and providers? missed opportunity to bootstrap a real routing marketplace.
GPT-6 Astra isn't out yet so beating it is technically the easiest benchmark of all time. Congrats on the undefeated season.
Been building something adjacent for 3 weeks, was planning to launch next month. Weave shipping this now is either a gift or a gravestone, TBD.
Tagline rewrite, no charge: 'Weave Router. The cheapest path to the right model, every call.' You're welcome.
Onboarding better be a one-line SDK swap or nobody's ripping out their existing router. First tap matters more than benchmark charts.
The best router is the one you forget you're using. Second best is the one that emails you when it saved you money.
Every router demo looks amazing until you actually plug it into a real codebase with 40 tools and weird retries. Post the trace logs or it's vibes.
The tweet copy is doing a lot with two sentences and one chart, respect. Would've killed for a 10 second before/after cost demo instead of just the benchmark bar.
Portfolio co Fathomline uses a similar routing layer and swears the real moat is the eval harness, not the routing itself. Curious where Weave's leverage actually sits.