OpenRouter
Type: app · Vendor: OpenRouter · Language: API (REST, OpenAI-compatible) · License: proprietary · Status: active · Status in practice: mature
Hosted LLM aggregator that exposes a single OpenAI-compatible endpoint over hundreds of models from many providers, with built-in provider routing, automatic fallback, price-weighted load balancing, and an Auto Router that picks the best model per prompt.
Description. OpenRouter is a commercial LLM aggregation service. Applications point at a single OpenAI-compatible endpoint and select a model by string identifier; OpenRouter handles authentication to the underlying provider, load-balances across providers prioritising price, falls back to the next provider on error, and offers an Auto Router (openrouter/auto, powered by NotDiamond) that selects the best model per prompt. Callers can customise routing via a `provider` object (e.g. explicit `order` of provider slugs). It is not a developer SDK in the build-your-agent sense — it is a hosted service end-user applications consume.
Agent loop shape. Hosted routing gateway. The application issues an OpenAI-style chat completion request to OpenRouter's endpoint with a model identifier (a specific model or a routing variant such as openrouter/auto). OpenRouter dispatches to one of the backing providers — by default load-balancing across the top providers weighted by inverse-square of price — applies fallback on errors, accounts cost, and streams the response back. Stateless per request; routing policy lives in the model slug, the optional `provider` object, and account settings.
Primary use cases
- unified access to many model providers from one endpoint
- provider fallback and price-weighted load distribution
- cost-aware automatic model selection via Auto Router
- OpenAI-compatible drop-in for multi-provider apps
Key concepts
- Unified OpenAI-compatible endpoint (docs) — A single API URL fronting hundreds of models from many providers; clients keep using the OpenAI SDK and just change the base URL and model slug.
- Price-weighted load balancing (docs) — Default routing across stable providers selects one weighted by inverse square of price, favouring cheaper backends without sacrificing uptime.
- Automatic provider fallback (docs) — If a provider errors or times out, OpenRouter transparently retries the next provider in the routing list.
- Provider routing preferences (docs) — Callers can pass a `provider` object (e.g. `order: ["anthropic", "openai"]`) to constrain or reorder the provider list per request.
- Auto Router (openrouter/auto) → cost-aware-action-delegation (docs) — Meta-model slug powered by NotDiamond that picks the best underlying model per prompt, routing simpler prompts to efficient models and complex ones to capable models.
- Latest-alias model slugs (docs) — Slugs like `openai/gpt-latest` resolve to the current flagship of a provider so client code keeps pointing at the freshest model without redeploying.
Patterns this app implements —
- ★★Multi-Model Routing
Whole product is built around routing each request across a catalogue of models and providers behind one OpenAI-compatible endpoint.
- ★Cost-Aware Action Delegation
Default load balancing prioritises price; the Auto Router delegates model selection to a meta-model that routes simpler tasks to cheaper models — both let the caller delegate cost-aware model choice…