Skip to main content
The AI Gateway API provides model discovery, inference, embeddings, and image generation. Clients choose a request format and model while project policy constrains eligible deployments.

​
Models and deployments

A model describes capabilities and accepted inputs. A deployment identifies where that model is served, including its residency, precision, retention, and pricing metadata. The public catalog exposes customer-facing metadata. Authenticated discovery can reflect a project’s inference policy. An available model does not imply that every deployment or protocol is available for a particular request.

​
Request formats

Inference endpoints support provider-compatible formats, including OpenAI and Anthropic formats. The provider gateway forwards requests using a provider-specific path. The endpoint reference defines supported paths, parameters, and response handling. Image generation produces project uploads. The Projects API manages those stored assets. Embedding generation produces vectors; the Knowledge API stores and searches indexed vectors. A support application can request a model response directly or use model inference as part of an agent run. A gateway request and an agent run are different resources: the Execution API provides the orchestration history for the latter. A model deployment identifies where inference is served. It is distinct from an application deployment, which assigns a project release to an environment.

​
Policies and spending

Project inference policies restrict model choices within the account rules they inherit and the deployment’s serving restrictions. Account API keys that name no project follow the account policy. Deployments retaining the legacy EU-only default enforce serving restrictions alongside model rules. Changing a model default does not remove a project tier restriction. See Use models served outside the EU. Budgets and usage belong to the Billing and Usage API. Authentication and API key ownership belong to the Identity and Access API.

​
Get started

Generate a model response shows model discovery and a chat-completion request.

​
API references