Intelligent
model router for

coding agents

About half the cost per trial compared with GPT-6 Astra on two coding benchmarks.

  • 2.2–2.5× faster per trial
  • Terminal-Bench 4.0: 62.1% vs 60.6% solved
  • SWE-Atlas: 62.1% vs 66.1% solved

Solve rates compare Router with GPT-6 Astra over two attempts per task.

Loading router example…refactor the auth module: DeepSeek, score 91, cost $$. debug a flaky integration test: Claude, score 96, cost $$$. summarize this pull request: Llama, score 74, cost $. triage 40 support tickets: Gemini, score 82, cost $. migrate the config to YAML: Kimi, score 79, cost $$. draft the release notes: GPT-5, score 88, cost $. explain this stack trace: Claude, score 90, cost $$.

Weave reads every turn your agent takes and sends it to the cheapest model that’ll get it right — saving the frontier ones for when it counts.

Save the frontier modelfor the turns that count.

Trained on 10× the sessions of 1.0.

A brand new session analysis model.

Benchmarks

Comparing Weave Router to Astra

Same Codex harness, same tasks, two attempts each — Router and GPT-6 Astra paired on the same day and package.

Compared with GPT-6 Astra

62.1%tasks solved in two attemptsAstra solved 60.6%. One additional task solved out of 66.
Tasks solved · pass@2Higher is better
  1. Weave Router62.1%
  2. GPT-6 Astra60.6%
  3. OpenRouter Auto25.0%

Terminal-Bench 4.0

62.1%tasks solved in two attemptsAstra solved 66.1%. Five fewer tasks solved out of 124.
Tasks solved · pass@2Higher is better
  1. Weave Router62.1%
  2. GPT-6 Astra66.1%
  3. OpenRouter Auto54.0%

SWE-Atlas

Cost and time are per trial. Quality is pass@2: a task counts if either of two attempts passed. Router and Astra are paired; OpenRouter is a separate run on the same tasks and graders. Charts share a scale within each metric.

Try Weave Code Max free for your first month with $10 in credits for $50 in usage, then pay $10/month.

How production token savings are measured

Weave’s published production results report 40% average token-spend savings with Router. This historical aggregate describes the production usage observed; evaluate your own workload against a fixed comparison model.

In production, Router reporting estimates savings against a fixed comparison model you choose. It adjusts for each harness's token-work ratio, cache effects and task success where that signal exists. Subscription-served requests carry no per-request API charge and show no dollar savings, and ineligible or unmeasured calls are reported at zero. Data can be exported for finance review.

Read the savings methodology before comparing the report with your invoice.

How it works

Four decisions behind each coding request.

From the blogWeave Router 2.0: Astra quality at half the cost
  • How hard is this turn?

    A complexity classifier trained on 10× the sessions of 1.0 scores it in single-digit milliseconds.

  • Will switching cost more than it saves?

    Cache-aware. It only routes down when the saving beats the cost of rebuilding the cache.

  • Is the cheap model actually finishing?

    An escalation classifier watches the task. Stall, loop, or miss → back to a frontier model.

  • How is usage funded?

    Uses each provider key’s Before plans or After plans funding priority.

Easy prompt · Oct 29$9.26 less

push and post replies

Sent to

Claude Opus 5.5

$10.80

Router

Claude Haiku 5.5

$1.54

One example turn, not a benchmark. Benchmark results are above.

Integration

Your tools stay the same.

Configure your coding client to send requests through Router. Add the provider keys you want to use.

npx @weave-os/router

Installers are available for Claude Code, Codex, Copilot CLI, OpenCode and Pi. Cursor uses manual configuration. In Claude Code 2.1.242 and later, Weave Auto appears in the native model picker, so engineers keep their workflow. Setup for each client, including the separate local flow for Claude Desktop, is in the Router documentation.

Models

The Weave model pool.

Every turn goes to the cheapest model in the pool that clears your quality bar.

  • Claude

  • GPT

  • DeepSeek

  • GLM

  • Gemini

All comparisons

Managed coding routing or a self-hosted gateway

This comparison is Weave’s, based on LiteLLM’s public product page as of November 2026. LiteLLM is an open-source AI gateway and proxy for platform teams.

LiteLLM also documents classifiers, cache-aware switching, escalation and subscription support. Weave's addition is an operated coding policy and client setup. Teams can adopt that combination without building and maintaining the gateway themselves. Its savings report uses your chosen comparison model with cache and task-success adjustments, which gives an engineering leader a coding-specific baseline. The two can coexist, with LiteLLM governing application traffic and Router serving coding clients.

NeedBetter starting pointReason
One gateway for all internal LLM applicationsLiteLLMProvider coverage, budgets, rate limits and caching for traffic your services generate
Lower model cost for Claude Code, Codex or Copilot CLI sessionsWeave RouterPer-request selection with cache-switch and escalation built for coding sessions
Self-hosted gateway on your own infrastructureLiteLLM, or the open-source Weave Router releaseLiteLLM is self-hosted by design; Router's public release runs a different policy from managed Router
Connect routing spend to merged code and reviewWeave Router with Weave analyticsToken Intelligence relates AI spend to estimated expert-effort output

Find your Router plan.

Choose a coding subscription or deploy Router in your own infrastructure.

Code Max

Perfect for indie developers.

A coding subscription with intelligent routing across open-source models only.

$10/month

Explore Code Max

Enterprise

For teams that need Router in their own infrastructure.

Run intelligent model routing on-prem with hands-on deployment support.

Custom

Request pricing

Funding, fees and self-hosting

How do the included allowances work?

Router plans are billed per seat and separately from Weave’s analytics platform plans. Included allowances belong to each seat and apply across six-hour, weekly and monthly windows. They are not pooled or transferable credits.

Code Max is free for the first month with $10 in credits for $50 in usage, then costs $10/month.

How are provider keys funded?

Bring-your-own-key (BYOK) funding follows each provider key’s Before plans or After plans setting; it is not always the last funding source.

Code Max keeps its open-source model restriction under every funding source.

What fees apply to credits and BYOK requests?

Purchased credits form a shared organization balance, and top-ups add 5.5% with a $0.80 minimum. BYOK requests carry a 5% platform fee drawn from the Weave credit balance. Enabling paid fallback does not authorize automatic recharge. The full rules are in Router billing.

Does installing Router also enable session analytics?

Router is independent from the Weave telemetry plugin. Router changes which model serves a request; the plugin captures sessions for Weave's analytics, cost and AI Skills reports. Installing Router does not turn on session analytics, and analytics customers do not need Router.

How does the open-source release differ from managed Router?

The public weave-os/router repository is Apache-2.0 licensed from router-v0.2.24 onward; earlier releases keep their original licenses. The public build ships a frozen HMM routing policy that uses Google's embedding API by default. That policy differs from the one behind the managed service, so test self-hosted results on your own workload instead of assuming the managed figures carry over.

Astra quality at half the cost.

Code Max is free for the first month with $10 in credits for $50 in usage, then $10 per seat per month. Router is billed separately from Weave analytics. A public Apache-2.0 release is available to self-host.