OpenAI API changelog

API separates traffic-spike errors from model overload errors

Traffic that ramps too quickly can now return a 429 error with code slow_down, and temporary model overload returns a 503 with code server_is_overloaded. Both may include Retry-After.

Who is affected:

API clients with retry logic keyed on status codes.

What to do:

Update retry logic to honor Retry-After and handle slow_down and server_is_overloaded.

Extended 24-hour prompt caching becomes the default

For organizations without Zero Data Retention, prompt_cache_retention now defaults to 24h instead of in_memory on the Responses, Chat Completions and Batch endpoints.

Who is affected:

API organizations without Zero Data Retention that never set prompt_cache_retention.

What to do:

Set prompt_cache_retention explicitly if you need in-memory only caching.