OpenAI's 12-hour prompt cache option pays off only after repeated rewrites

OpenAI's Responses API now lets developers opt into a 12-hour minimum cache lifetime instead of the default 30 minutes. That longer option doubles the write price, so on GPT-6.2 Astra a 12-hour write costs $20 per million tokens versus $10 for the shorter window. The analysis suggests the extra cost is recovered only when a stable prompt prefix would otherwise expire and need rewriting at least twice. Frequently changing prefixes or workloads that reuse context within 30 minutes can lose money under the 12-hour setting. The author recommends measuring cached-token usage per step before switching the default.
This is an AI-generated summary. ShortSingh links to the original source for the complete article.

Discussion (0)
Log in to join the discussion and vote.
Log in