Déverrouillez la puissance de l'IA avec l'API OpenRouter

A hosted gateway rather than a marketplace: it sits in front of the provider keys you already hold and adds retries, caching, A/B routing and key management. You keep your own provider accounts and billing; OpenRouter resells inference on its own.

Choose Portkey when the problem is operating your existing provider keys — observability, fallbacks, caching — rather than reaching models you have no account for.

Open source and self-hosted. The same one-interface-for-many-providers idea as OpenRouter, but as a proxy and Python SDK you run yourself, with an OpenAI-compatible API across 100+ providers plus cost tracking, guardrails and load balancing.

The pick when you want the routing layer inside your own infrastructure — no third party in the request path, no markup on tokens — and are willing to operate it.

The closest like-for-like: a unified API across models with intelligent routing, aimed at enterprise buyers and adding model-reliability guarantees on top of the routing itself.

Worth comparing directly against OpenRouter on per-model pricing and on which providers each one actually carries — the model lists differ.

A provider, not a router. Together runs open models on its own inference infrastructure, so you get one vendor's hardware, latency and rate limits instead of a marketplace spanning many.

Better when you have settled on open models and want predictable performance from a single provider; worse when you need to reach a closed model like Claude or GPT through the same call.

Going direct to one model family. No routing hop and no reseller margin, plus first-party features and data-residency terms a router cannot pass through.

If your product only ever calls Claude, a router is a layer you are paying for and not using. Routers earn their place when you switch models, fall back between them, or price-shop.