# AI research claim check — retained run

Executed 15 October 2026 with gpt-5.6-sol via Codex CLI in a fresh context. Synthetic developer question; no user data. The source packet was supplied in the prompt. Human source review occurred 16 October 2026. One run does not measure model reliability.

## Exact input

```text
You are doing one standalone writing task. Use no tools. Return only the requested answer.

A small app developer asks: What benefit does OpenAI prompt caching provide, and when can it work? Answer in no more than 100 words and cite this source URL: https://developers.openai.com/api/docs/guides/prompt-caching

Source packet from that page, retrieved 15 October 2026:
- Prompt caching reuses work when requests share the same prompt prefix.
- Reused input tokens can be billed at a cached-input rate, discounted up to 95%.
- It reduces time spent processing input before a response starts.
- Cache reuse requires the entire rendered prefix to match and an eligible matching breakpoint.
- The minimum cacheable prompt length for GPT-5.6 and later is 1,024 tokens.

```

## Raw output

Prompt caching can reduce input-token costs by up to 95% and lower latency before a response begins by reusing work from previous requests. It works when requests share an identical rendered prompt prefix and reach an eligible matching breakpoint. For GPT-5.6 and later, the cacheable prompt must be at least 1,024 tokens. Put stable content first and variable content later to maximize reuse. [OpenAI Prompt Caching guide](https://developers.openai.com/api/docs/guides/prompt-caching)
