A longer cache can help when the same context returns hours later. Its higher write charge needs enough reuse to justify the choice.
What changed in the API
OpenAI’s October 20 changelog adds a 12-hour prompt-cache option. The current guide describes supported Responses API models and the lifetime setting.
For GPT-5.6 and later, the guide lists default 30-minute writes at 1.25 times ordinary input cost. The 12-hour write rate is 2.5 times the ordinary input cost.
Cached reads use a different rate. The guide lists 0.1 times ordinary input for most of these models and 0.05 times for GPT-6.1 Sol.
Use an explicit reuse scenario
Consider three requests with the same eligible prefix. Assume the shorter policy requires a new write each time, while the longer policy permits two cached reads.
At a 0.1 read multiplier, the shorter case costs 3.75 times the prefix’s ordinary input cost. The longer case costs 2.7 times that cost: 2.5 plus 0.1 plus 0.1.
This illustrative difference concerns only the reused prefix. It excludes output, uncached input, tools, and retries. It also assumes the stated writes and hits actually occur.
If the shorter policy retains the prefix and produces hits, that comparison changes. Measure the real reuse pattern before treating the example as a savings estimate.
Read the usage fields without counting writes twice
The guide separates cached reads from cache-write tokens and provides lifetime-specific write counts. Use the matching rate for each category.
I would retain request-level usage for a representative day. Group identical eligible prefixes by their reuse gaps, then compare policies against those observed calls.
- Record the requested model and cache policy.
- Separate ordinary input, writes,s and cached reads.
- Include requests that changed the prefix.
- Keep output and external tool charges in the total.
- Compare accepted work and correction time alongside spending.
Choose a policy for the workload
A frequently reused prefix can behave differently from reference material read only a few times across a day. Keep those workload classes separate.
API prompt caching is distinct from ChatGPT conversation history. A longer lifetime does not mean permanent storage or a promise that every later request will hit the cache.
Our agent cost-control guide covers failed attempts and budgets. The Claude inline-tools article examines a different cache-related interface.
Read the sources and choose your next check
How many eligible reuses would your stable prefix receive during a typical working day?
Checked October 22, 2026. This article is based on source review and analysis. Suggested tests have not been performed by MustHave.ai.
