API conventions glossary

Links on this page to features marked Preview describe designs that are not available yet.

Terms shared by every endpoint. Shorter definitions of the core API terms are on the main Glossary.

Extra information attached to a reply, such as a citation. See citations.

Replacing old turns with a summary so a conversation fits the context window.

Server-side clearing of old tool results and compaction of long conversations. See context management.

Paging with after_id and before_id, using first_id, last_id and has_more from each page.

The specific, stable reason for an error. Branch detailed handling on it. See error codes.

The one shape every error uses: error.type, error.code, error.message, error.param, error.request_id. See errors.

The broad category of an error, fixed by HTTP status. Decide whether to retry from it.

An opt-in context window of up to 1,048,576 tokens on supported models. See extended context.

A value you send so a retried create does not happen twice. Kept for 24 hours.

A comment line sent during long streams so connections are not closed as idle. See stream format.

A reply in which the model declines; finish_reason is refusal and message.refusal explains. See edge cases.

The request field that asks for JSON output, optionally matching a schema. See json output.

Seconds to wait before retrying, on 429 and 503 responses. See rate limit headers.

The header that turns on preview features by flag. See versioning.

A response header warning that a version or model you use is scheduled to go away. See deprecation policy.

The header that pins an API version by date. See versioning.

Was this page helpful?