OpenAI Adds 12-Hour Prompt Cache for GPT Models
The new 12-hour cache option allows developers to sustain shared context across requests for longer durations at a set premium.
Updated on Oct. 21, 2026 in Artificial Intelligence
Live Poll
Do you plan to increase your use of AI tools despite rising costs for cached data?
OpenAI has introduced a 12-hour minimum prompt-cache lifetime for its GPT-5.6 and newer models. This update joins an existing 30-minute setting, allowing developers to choose cache durations based on their specific context-reuse needs.
Why it matters
The change allows developers to maintain shared context across requests made hours apart, reducing repetitive processing. It marks a shift toward tiered cost structures that prioritize context-heavy workflows over standard input-token billing.
The 12-hour prompt cache option costs 2.5 times the standard input-token rate, while the 30-minute option remains at 1.25 times. Subsequent cache reads cost one-tenth of the standard input rate, regardless of the chosen lifetime setting.
The players
OpenAI
An AI research and deployment company that produces the GPT series of large language models.
The details
Developers manage these lifetimes via the Responses API by selecting the desired duration. This system caches prompt context—the preceding text or instructions provided to the model—allowing the engine to reference stored data rather than re-processing it. While the setting acts as a minimum guarantee, the system may store context for longer durations as required by infrastructure load.
Timeline
October 20, 2026: OpenAI introduced the 12-hour minimum prompt-cache setting.
The Tech Race
This update expands the caching capabilities of the OpenAI Responses API, moving beyond the original 30-minute threshold to accommodate longer-running application workflows. It reflects a growing industry focus on lowering latency for complex, multi-turn AI interactions.
Developers using GPT-5.6 or later can now optimize costs for applications that require maintaining context over several hours. Meanwhile, users of Codex and ChatGPT Work gain indefinite context retention, removing the need for manual cache management in those specific environments.
The takeaway
Developers should evaluate the cost-to-latency trade-offs for their specific workflows, as the 2.5x write rate for 12-hour caching is a premium over the 30-minute option. Monitor subsequent API documentation for any updates to cache retention limits beyond the guaranteed minimums.
Further reading
For broader trends in model efficiency, visit the Artificial Intelligence section.
Live Poll
Do you plan to increase your use of AI tools despite rising costs for cached data?
