On October 20, 2026, OpenAI introduced a 12-hour prompt-cache option. Current documentation prices the 12-hour write at 2.5 times ordinary input: that is the total write rate, not an extra surcharge. The longer lifetime can help repeated work hours apart, but it does not automatically save money. Check support for the selected model and endpoint.
For GPT-5.6 and later, the documentation distinguishes cache writes from reads. Twelve hours is a minimum lifetime, not a promise of deletion at that time.
Prompt caching can reduce the cost of reused input, but repeating a request does not guarantee a saving. The model, matching prefix, cache lifetime and actual usage fields matter.
Measure reads and writes separately
Keep stable instructions and reference material before the changing task. Then inspect usage.input_tokens_details.cached_tokens and, where reported, cache_write_tokens. A cache write and a cache read are different billing events; current models can charge extra for writes. Check the selected model’s current pricing rather than applying an older model’s rule.
For example, compare ten document checks that share one reference guide. Record ordinary input, cache writes, cache reads and output for the whole group. One fast second response is not a cost comparison.
Check the useful result
Verify that the changing document was actually included and that cached material was still current. Caching does not make output tokens free or increase the API’s rate limits. Keep the baseline and cached totals in the same currency and pricing tier.