OpenAI Extends GPT-5.6 Prompt Caching to 30 Minutes
The builder guide adds deterministic cache breakpoints and same-engine routing hints for repeated agent context, while leaving real-world savings to workload testing.
OpenAI published a GPT-5.6 builder guide on August 13 that documents a prompt-caching change for developers running repeated or long-context agent requests. The official builder guide says the model family's prompt-cache lifetime has been extended to a minimum of 30 minutes, cache breakpoints can be set deterministically within a model's context, and an appropriate cache key can improve the likelihood that related requests reach the same inference engine.
The change matters to teams whose agents repeatedly carry large instructions, tool definitions or shared workspace context. A longer reuse window and explicit breakpoints could reduce the amount of unchanged context that must be processed again, but OpenAI does not establish a universal saving. Its routing language describes a higher likelihood rather than a guarantee, and the published evidence is first-party.
What changed in GPT-5.6 prompt caching
OpenAI says the 30-minute minimum applies across the GPT-5.6 family. Deterministic breakpoints give developers a defined place inside the context for cache handling instead of leaving the entire reuse decision implicit. The guide also points developers to `prompt_cache_key` as a way to increase the chance that a request lands on the inference engine that previously handled the same prefix.
Those pieces address different parts of the same problem. The longer lifetime keeps a reusable entry relevant for more of a working session. A breakpoint identifies the context boundary that a team intends to reuse. The key influences routing. None of them, on the evidence provided, guarantees a cache hit or a particular latency or cost reduction.
Where caching fits in OpenAI's agent stack
The guide presents prompt caching alongside retained reasoning, native context compaction, multi-agent orchestration and programmatic tool calling. OpenAI's stated design goal is to preserve model tokens for judgment while moving deterministic filtering, aggregation and tool coordination into code. The cache update is narrower: it focuses on reusing context that has already been processed.
That distinction matters when evaluating an agent system. Better caching can improve the economics of repeated context without changing the model's underlying intelligence. Retained reasoning and compaction affect how work continues across a longer task, while multi-agent and programmatic-tool features change how work is divided. Teams should measure these effects separately rather than attribute every improvement to the model itself.
What teams still need to verify
OpenAI includes selected startup experiences and company-run results in the guide, but those examples are not independent evaluations and do not establish the outcome for another prompt shape, tool set or traffic pattern. The public page also does not provide a workload-neutral comparison of cache-hit rates before and after the change.
For production evaluation, the useful checks are observed cache reuse, cached versus uncached input, end-to-end latency and total spend on the same workload. Because OpenAI says the cache key only increases the likelihood of same-engine routing, teams should treat cache performance as something to measure rather than assume. The next stronger evidence would be reproducible tests across long-running agents with different amounts of shared and request-specific context.
Status
Confirmed. Internal confidence is medium because OpenAI's official guide documents the cache changes, while the performance implications and customer examples remain first-party and lack independent workload testing.
Sources
Update note: Last reviewed 2026-08-14. We will revise this post if OpenAI changes the cache controls or independent tests establish their effect on real agent workloads.
Sources
- OpenAI — The builder's guide to GPT-5.6 — official
Drafted with AI assistance from source briefs; reviewed for citation completeness and label accuracy.