C04 — Prompt caching

• Publisher: OpenAI
• Source: Prompt caching
• Published: Not stated / not established
• Updated: Not stated / not established
• Accessed: 2026-10-22
• Type: Official implementation documentation
• Evidence: opened primary page; relevant content read.

Paraphrased finding

Caching reuses an unchanged prefix. GPT-5.6+ has a 1,024-visible-token minimum and separately priced writes. Verify actual read/write counters.

Limits

Requires reuse while cache is valid; cache read discount is not a whole-application saving; model and retention settings differ.