RouterBench

•  the coding agent  •

Don't pick a model.
Ship.

RouterBench classifies every request and routes it to the best model — ours included. Watch the router work in your own terminal, editor or desktop.

secure · specialized · enterprise-grade

Ask anything… "Fix broken tests"

Build · routerbench/auto RouterBench

~/project  tab agents  ctrl+p commands

v1.18.3

Get RouterBench on your machine.

Sign in once with your browser and use RouterBench everywhere — the desktop app, your terminal, and VS Code all share one account and one subscription.

free to start · every model · no card required

OpenAIAnthropicGoogleGroqMoonshotZ.aiSarvamMistralxAIMeta LlamaQwenPerplexityCerebrasFireworksTogetherNovitaAINebius130+ models25+ providersone apirouterbench/autoOpenAIAnthropicGoogleGroqMoonshotZ.aiSarvamMistralxAIMeta LlamaQwenPerplexityCerebrasFireworksTogetherNovitaAINebius130+ models25+ providersone apirouterbench/auto

the problem

Choosing AI became harder than building software.

gpt-5claudegeminiqwenkimi

scanning the market · every model · live

too many models

01

A new frontier model every month.

GPT, Claude, Gemini, Llama, Qwen, Kimi — each better at something. Nobody can track which wins where.

prices shift weekly

02

Costs move under your feet.

Token prices drop, tiers change, caching rules differ. Yesterday's cheapest model is today's overspend.

benchmarks go stale

03

Leaderboards lie by launch day.

Static benchmarks can't tell you how a model performs on your tasks, on your traffic, right now.

providers go down

04

Outages don't announce themselves.

Rate limits, degradations and downtime hit mid-request. Without failover, your product wears the outage.

RouterBench makes the decision for you — on every request.

•  introducing RouterBench  •

Everything routed.
Nothing to choose.

Send model: routerbench/auto. Each request is classified, scored against the live AI Credit Score™, guarded, and routed to the best model — with automatic failover when a provider degrades.

how a request flowsmodel: routerbench/auto
RouterBench cliterminal agentvs codeextensiondesktopmac · win · linuxplaygroundbrowseryour appopenai sdkRouterBench smart routerthe intelligence layer01classifytask · domain · complexity · sensitive02scoreai credit score™ · quality × cost × latency03policysensitive→on-prem · escalate · fallback04selectavail 50 · speed 20 · cost 20 · latency 10claude-opushosted · reasoninggpt-5hosted · codegemini-flashhosted · fast · cheaprouterbench-devour model · first-partyqwen-localon-prem · sensitive
5surfaces
130+models
25+providers
1api · one bill

•  everywhere you build  •

One intelligence layer. Every surface.

Every surface shares one account and routes through routerbench/auto — the right model handles every task, wherever you work.

•  ai credit score™  •

A live rating for every AI model.

RouterBench continuously scores every model on accuracy, reliability, cost efficiency, latency and real-world performance — per task. Route on trusted, up-to-date data instead of static benchmarks or gut feel.

  • live Updated from real traffic — the score moves when models move
  • per-task The best model for code isn't the best for extraction
  • unbiased Every provider scored by the same yardstick
live · success rate
claude-sonnet-4-5
100
claude-sonnet-4-6
89
text-embedding-3-small
100
claude-haiku-4-5
95

•  the platform  •

Your AI stack, unleashed.

01 —

AI Decision Engine

Selects the best model for every request on task, quality, speed and cost. No manual comparison, no switching.

02 —

AI Credit Score™

A live rating for every model — accuracy, reliability, cost, latency, real-world performance. Data, not benchmarks.

03 —

Intelligent Routing

Routes each request to the optimal model; switches providers automatically during outages and slowdowns.

04 —

AI Cost Optimizer

Continuously routes to the most cost-effective model without compromising quality. Spend less, ship more.

05 —

Guardrails & PII

Detects and redacts PII/PHI/secrets before any cloud hop. Sensitive prompts can be forced on-prem, fail-closed.

06 —

Observability

Logs, traces, per-request USD cost and latency analytics across every model. OpenTelemetry-native.

chat · embeddings · audio · realtime voice · image · music · video — one endpoint

models

130+

across every major provider

providers

25+

with automatic failover

api

1

OpenAI-compatible endpoint

markup on byok

$0

your keys, provider rates

•  the model lab  •

We don't just route models. We build them.

The intelligence layer sees which model wins every task — and that intelligence trains our own. RouterBench is a model company and a provider, not just a switchboard.

01

RouterBench models

Models of our own.

We train and serve our own models — built on real developer workflows and the routing intelligence flowing through the layer every day.

02

first-party provider

RouterBench is a provider too.

RouterBench models run on our own infrastructure, served through the same API — and compete for every request under the same AI Credit Score™ as everyone else.

03

specialized > general

Purpose-built, not general.

Not another everything-model. Specialized models for developer work — tuned for what real-world engineering tasks actually demand, privacy-preserving by design.

04

industry-specific

Post-trained for your domain.

Learn more

We post-train open foundations into industry-specific models — our fine-tuning, our alignment, our domain evaluation. Benchmarked on your real tasks and first-class in the router.

•  quickstart  •

Three lines to switch.

RouterBench speaks the OpenAI protocol — a one-line change gives you smart routing, observability and cost control for free.

  • drop-in Your existing OpenAI SDK code keeps working
  • resilient Failover, retries and caching built in
  • transparent Per-request cost in USD on every response
app.ts
import OpenAI from "openai";

const client = new OpenAI({
  baseURL: "https://api.routerbench.com/v1", // ① point at RouterBench
  apiKey: process.env.ROUTERBENCH_API_KEY,
});

const res = await client.chat.completions.create({
  model: "routerbench/auto", // ② the router decides
  messages: [{ role: "user", content: "Hello" }],
}); // ③ best model answers

•  enterprise  •

Governed. Private. Deployed with you.

Give every team access to every model — without giving up control.

01identity

SSO · RBAC · audit

OIDC SSO (Entra, Okta), org roles, and an append-only audit trail for every action.

02deploy

On-prem & air-gapped

Managed cloud, your VPC, or fully self-hosted with your own storage and models.

03data

Sensitivity firewall

PII/PHI/secrets tokenized before egress; sensitive prompts forced on-prem, fail-closed.

04people

Forward-Deployed Engineers

Our engineers design routing, guardrails and governance with your team — and keep it running.

•  faq  •

Frequently asked questions.

Start free, then subscribe to Pro or Max for a flat monthly plan with generous rolling 5-hour and weekly usage windows — no per-token bills, the plan covers every model. Enterprise is custom, billed per contract. Prefer pay-as-you-go? Top up credits and spend only when you route.

Yes — run it as managed cloud, inside your own VPC, or fully self-hosted and air-gapped with your own storage and models.

130+ models across OpenAI, Anthropic, Google, Groq, Moonshot, Z.ai, Sarvam and more — plus your own custom, fine-tuned or self-hosted endpoints.

Yes. Point the OpenAI SDK at api.routerbench.com/v1 and change one string — your existing apps, agents and tools keep working, and you get routing, observability and guardrails for free.

Install the CLI (npm i -g @routerbench.com/cli), the VS Code extension, or the desktop app. All of them route through routerbench/auto, so the right model handles every task.

The sensitivity firewall detects and tokenizes PII/PHI/secrets before any cloud hop; policy can force sensitive prompts to on-prem models so raw data never leaves your network.

Register any OpenAI/Anthropic-compatible endpoint (Ollama, vLLM, LM Studio, TGI) as a first-class provider, and the router picks it whenever it fits the task.

Yes. RouterBench is a model company and a first-party provider — we train models on real developer workflows and routing intelligence, serve them on our own infrastructure, and let them compete for every request under the same AI Credit Score™ as every other model.

can't find your answer? get in touch

Stop choosing.
Start building.

Point one endpoint at RouterBench — or install the coding agent — and ship your first request in minutes.