One request shape, one streaming contract, one error model, one usage record — across OpenAI, Anthropic, Google, Meta, Mistral, DeepSeek and the rest of the catalog.
Provider APIs look similar until they do not: different message roles, different tool-call shapes, different ways to report token usage and different error codes for the same failure.
A unified LLM API normalises all of it. You write one integration, and models become configuration rather than code.
The gateway translates between the unified contract and each vendor API in both directions, including streaming.
Because the model is a request parameter, moving traffic from one provider to another is a configuration change. Teams use that to cut cost on high-volume paths, move long-context work to a cheaper model, or shift load when a provider degrades.
New providers ship behind the same contract, so adopting a model that did not exist when your feature launched does not require a new client library or a new retry strategy.
Short answers about the gateway, the providers and how billing works.
Yes. Server-sent events are normalised so a stream looks the same regardless of the provider that produced it.
Embeddings are available through the same endpoint family and follow the same OpenAI-compatible request and response shape.
Yes. Extra parameters can be passed through per request, and the gateway ignores or maps what a given provider cannot accept.
Failures are mapped to a consistent error taxonomy with the provider reason preserved, so your retry logic does not need to know every vendor code.
Sign in with Google, open the playground and price a request against the whole catalog before you change a line of production code.
