Gateway
Requests & turns
One request is one metered turn. The gateway holds no conversation state.
Turns
A chat is a messages array that grows over time, and your client sends the full array on every request. A ten-turn conversation is ten requests, each metered on the tokens it carried in and out.
Agent loops
When the model returns a tool call, your client appends the result and sends the transcript again. Each round trip is a turn. /v1/responses accepts the same loop in the Responses shape and is just as stateless.
Switching models mid-conversation
Name a different id on the next request and send the same transcript. Nothing ties a conversation to the model that started it, so you can run a cheap model until an answer disappoints and then resend the same messages to a frontier id. Capabilities belong to the id. Read caps on the new model before you send tools or images.
Streaming, parameters, time bounds
Streaming, parameters and time bounds are defined once, in the API reference.