OpenRouter

OpenRouter is an LLM API gateway for developers: one key and one endpoint for models from many AI labs, paid as you go with a platform fee on credits.

OpenRouter lets an application call language models from many labs through a single API, without opening an account with each provider. A team buys credits once and names the model it wants in every request. Each model stays the property of the lab that trained it. OpenRouter, Inc. handles the routing, billing and controls around those models from New York, and has since early 2023. The API accepts the OpenAI chat completions format, the Anthropic Messages shape and the Responses API. It standardizes tool calling and structured outputs across providers, while image input, audio input and reasoning effort work on the models that support them. Auto-routing and preferred vendor selection decide which provider serves a model, and Zero Data Retention routing can be enforced for an account, a model group or one request. Prompt caching, a Batch API and an Embeddings API share the same key. Admins get budgets and spend controls, activity logs, a management API key and Bring Your Own Key for provider accounts they already hold. The service is built for indie developers, AI-native startups, growing product teams and enterprises that put language models into their own software. It suits a team that expects to change models often, or one that wants a single bill in place of several provider invoices. Buying works without a subscription: customers add credits, pay each model's listed price, and OpenRouter takes a platform fee on every credit purchase. Companies with residency rules move up to Business for EU or US in-region routing. Enterprise adds SSO, contractual SLAs and invoicing under a negotiated agreement. A common job is moving a product between models. Code written for the OpenAI API format points at OpenRouter's endpoint and names the model in each request. Product teams mix text, image, video and speech models in one feature, while data teams push offline workloads through the Batch API. Engineers try candidates in Chat and check Benchmarks, which measures models and providers on the platform itself, before they commit. It fits best where the choice of model keeps changing and nobody wants a new integration each time it does.