Weave Router Review (2026): Pricing, Features & Honest Verdict

Weave Router is a source-available model router that sits between coding agents like Claude Code and Codex and their providers, sending each request to the cheapest capable model. Verdict: candid benchmarks and a fair fee, but marketing runs ahead of the docs. Best for: teams with real agent bills. Price: 5% of routed spend, free to self-host. Rating: 7.5/10

What is Weave Router?

We found Weave Router at number one on Product Hunt for 16 September 2026, launched as “Weave Router 2.0” with the tagline “Subscription aware coding agent router.” It is a proxy. You point Claude Code, Codex, opencode, pi or Cursor at it instead of at Anthropic or OpenAI, and it decides which model should handle each request. Easy turns go to cheap models such as DeepSeek, Kimi or GLM. Hard ones stay on a frontier model. Your tools do not know anything changed.

Weave (Workweave, Inc.) is the company behind it, and its main product is an engineering analytics platform that measures how much work engineers and their AI agents get done. Its homepage claims more than 500 engineering organizations and SOC 2 Type II certification. Routing is a newer side of the business. The GitHub repository was created on 27 April 2026 and had 4,456 stars, 122 forks and a push on the day we checked, 18 September. Version 2.0 shipped on 9 September with a retrained classifier, cache-aware switching, an escalation path and routing across more than one subscription.

What Are Weave Router’s Key Features?

A classifier that scores every request

Each request is embedded by a small ONNX model running in-process and scored against clusters of past coding work. The README credits the approach to the Avengers-Pro routing paper (arXiv 2508.12631). Weave says the 2.0 classifier was trained on ten times the agent sessions of 1.0 and runs in single-digit milliseconds. The GitHub description says under 50 ms. Either way it is small next to a model call. A /v1/route endpoint returns the decision without calling a model, so you can inspect what it would do before you trust it.

Cache-aware switching and escalation

This is the part most routers skip. A coding session is a long run of requests sharing one context, and providers discount cached context heavily. Switch models mid-session and you throw the cache away and pay full price to rebuild it. Weave admits version 1.0 had exactly this problem. Version 2.0 tracks cache state per provider and per session and only routes down when the saving beats the rebuild cost. A second classifier watches the task, and if the cheap model stalls, loops or misses the point, the work moves back to a frontier model.

One-command install for the agents you already use

Run npx @weave-os/router and the installer asks which tool to configure, user or project scope, and wires the config file. For Claude Code the router key rides in a custom header, so your normal Anthropic login stays in place. For Codex it writes a managed block into config.toml and leaves everything else alone. You get /router-off, /router-on and /force-model commands to pin or bypass routing per session. Cursor support exists through the OpenAI base URL override, but the README labels it “early beta, performance may not be the best.”

Self-hosting with your own keys

Clone the repo, add an OpenRouter or provider key, and make full-setup boots Postgres and the router on localhost:8080 with a dashboard at /ui. The README is explicit that in this mode “prompts go from the router straight to the provider you configured, never to Weave.” Provider keys are encrypted at rest, traces go out over OTLP to Honeycomb, Datadog or Grafana, and every non-streaming response carries cost headers. A raw routing-decision export lets you pull the data into your own warehouse.

How Much Does Weave Router Cost?

Weave Router costs 5% of routed spend on the hosted version and nothing if you self-host. The launch post says “pricing for solo developers and startups is 5% of routed costs,” and the product page repeats “5% of routed spend.” Teams of 50 or more engineers are pointed at a forward-deployed engineer and a sales call.

Put numbers on it. Weave’s own calculator takes a team spending $12,000 a month on a frontier-heavy mix and estimates $5,640 after routing. Five percent of that is $282. If the estimate holds, you pay $282 to save $6,360. The calculator carries a fair caveat: actual savings depend on workload, prompt mix and provider pricing.

One thing needs fixing. The company’s main pricing page does not mention the 5% at all. It lists the analytics platform at $0 for Starter and $50 per engineer per month for Pro, then shows a separate box that reads “Reach out to our team about our prompt router” with a Contact Sales button. A solo developer landing there would assume the router is enterprise-only. It is not.

For comparison, OpenRouter charges a 5.5% platform fee on its Standard plan, and Not Diamond charges $0.05 per million tokens routed. Weave’s 5% is in line with both, and the self-host option is free.

Do Weave Router’s Benchmarks Hold Up?

Mostly, and the reason is that Weave published the caveats itself. The headline is “Astra quality at half the cost.” On Terminal-Bench 4.0, the router solved 62.1% of 66 tasks against 60.6% for GPT-6 Astra, which is one task. It took 20.5 minutes per trial against 44.1, and cost $5.22 against $10.03. On SWE-Atlas Codebase QnA, Astra won: 66.1% of 124 tasks against the router’s 62.1%, five tasks more. The router was again faster, 6.8 minutes against 16.7, and cheaper, $2.31 against $5.04.

So the accurate summary is a tie on one benchmark and a small loss on the other, at roughly half the cost and more than twice the speed. Weave says as much. The post states that the 95% confidence intervals overlap on both benchmarks, that scoring is pass@2, and that “k=2 is not a leaderboard submission” because Terminal-Bench requires five attempts. We rarely see a vendor print that under its own chart.

Three limits are worth knowing. The runs used the Codex harness only, and the bench README says Claude Code is “out of scope for now.” The router arm ran on a /beta lane, so check that the default lane you get matches what was tested. And the OpenRouter Auto comparison, 25.0% and 54.0%, was a separate run within three days, which Weave tells you to read “as context, not paired wins.” The harness lives in the repo under /bench, so you can rerun it.

What we could not support is the line in the site’s navigation calling it the “#1 Ranked Prompt Router In the World.” No ranking is cited on any page we read.

Is Weave Router Open Source, and What Does “Subscription Aware” Mean?

It is source-available, not open source. The licence is the Elastic License 2.0, copyright Workweave, Inc. You can read, modify and self-host the code. You cannot offer it to others as a hosted service or strip the licence notices. For an internal engineering team that restriction will never matter. For anyone planning to build a product on top of it, it matters a lot. GitHub’s API reports the licence as “Other,” which is what happens with ELv2.

The subscription claim needs care. The launch post says “Use your Claude Code subscription in Codex (and vice versa).” The install docs describe something narrower. Inside Claude Code, requests keep your normal Anthropic login for your Team, Pro or Max plan. Inside Codex, the native Codex models use your ChatGPT login. Every other model, the docs say, uses “their matching WorkWeave deployment or BYOK credentials.” As we read that, a Claude model chosen inside Codex is paid for through Weave or your own API key, not through your Claude subscription.

That is still useful. The router drains the flat-rate plan you already pay for before it spends metered tokens. But it is not the same as one subscription working everywhere. Before you point any flat-rate plan at a third-party proxy, read your provider’s terms for that plan. Vendors have changed those rules before.

Two smaller notes. On the hosted version your requests pass through Weave’s endpoint on their way to the provider, because that is what a proxy is, so teams with strict data rules should self-host. And the repo carried 130 open issues when we looked, which fits a project shipping this fast. The npm package has had 27 releases since 15 May, and it was renamed to @weave-os/router while the marketing site still prints the old @workweave/router name. The old name still works as an alias.

Who is Weave Router Best For?

Use Weave Router if your team runs Claude Code or Codex all day and the API bill has become a line item someone asks about. At $2,000 a month and up, a 40% to 50% cut pays for the 5% fee many times over, and the self-host path means security can say yes. Use it if you already want DeepSeek, Kimi or GLM for easy turns but do not want engineers choosing models by hand.

Skip Weave Router if you are one developer on a single flat-rate plan that you never exhaust. There is nothing to save. Skip it if Cursor is your main editor, since that path is early beta. Skip it if you need a general API gateway for a product backend. Weave says so itself: “Coding workloads only. Not a general API gateway.”

Best Weave Router Alternatives

OpenRouter is one API for more than 500 models with a 5.5% platform fee on the Standard plan. Bring your own keys and the first $25,000 of list-price inference each month carries no fee, then 5%. Its Auto router picks per request with no session or cache state. In Weave’s tests it scored 25.0% on Terminal-Bench and 54.0% on SWE-Atlas, in a separate run, so treat that as the vendor’s number.

LiteLLM is the open-source gateway that platform teams self-host, with 58.8k GitHub stars by its own count. It is free to run, with SSO, audit logs and an SLA sold through an enterprise plan. You get load balancing, fallbacks, budgets and spend tracking across more than 140 providers. It does not judge whether a cheap model can finish the task, so you write the routing rules yourself.

Not Diamond is a managed learned router priced at $0.05 per million tokens routed on pay-as-you-go, with multi-harness support and a savings dashboard. The catch is the button, which says “Apply for access.” Weave installs in one command today.

Final Verdict: Is Weave Router Worth It?

Yes, for teams with a real coding-agent bill. The design is right: route per session instead of per request, respect the prompt cache, and escalate when the cheap model struggles. Pricing is fair at 5%, the code is on GitHub under a licence most companies can accept, and the benchmark write-up is the most candid we have read from a vendor this year.

We rate it 7.5 out of 10. It loses points for a pricing page that hides the 5% behind Contact Sales, for a “#1 in the world” badge with no source, for a subscription pitch that reads bigger than the docs, and for benchmarks that so far cover Codex only. Run it in self-hosted mode for a week, read the savings line in your status bar, and compare it to your last invoice. The numbers will tell you quickly.