OpenAI extended the prompt cache lifetime: as of October 20, 2026, the prompt_cache_options.ttl parameter in the Responses API accepts the value "12h" instead of the standard "30m", and the cache for logged-in Codex and ChatGPT Work users is stored indefinitely, according to the Codex lead. The constant beginning of a request—instructions, tools, accumulated context—no longer needs to be re-warmed after every break: reading from the cache is up to 95% cheaper than standard input, although the higher write rate is only offset by regular access to the same beginning. For the market, this shifts caching from a technical optimization to a product pillar for working with models.

image
image

What happened

On October 20, 2026, OpenAI expanded prompt caching capabilities in the Responses API: the prompt_cache_options.ttl parameter now accepts the value "12h" alongside the standard "30m", and the unchanging beginning of a request—system instructions, tool definitions, accumulated context—can be kept in the cache for up to 12 hours instead of 30 minutes. OpenAI documentation charges a markup on the base input token price for cache writes: 1.25× for the 30-minute cache and 2.5× for the 12-hour cache. Reading from the cache is up to 95% cheaper than standard input: in an example from the post for the GPT-6.1 Sol model, cached reading costs $0.10 per 1 million tokens, compared to $2.00 for standard input and $2.50 for writing. Codex lead Tibo Sottiaux separately stated that for logged-in Codex and ChatGPT Work users, the cache is stored indefinitely: according to him, returning to a session after several hours guarantees a cache hit and does not spend limits on refilling it.

Context

Prompt cache operation is based on the structure of the request to the model: a significant part of each request is constant for a specific project or product because it is defined by system instructions, descriptions of available tools, and accumulated context, while only the ending part actually changes. When the service has already processed such a beginning, subsequent requests reuse the ready result instead of recalculating from scratch—the discount on reading from the cache is based on this difference. The announcement itself is not tied to models: there is no new architecture, training method, or benchmarks in it; the change lies in the area of serving and pricing. Its significance is purely engineering-economic, and it is documented in the official OpenAI changelog and guide as a working API parameter, not future plans.

Why this matters for the industry

Previously, a pause of more than 30 minutes reset the prompt cache and increased the cost of repeat requests—this primarily affected persistent agents and long working sessions. The 12-hour TTL and indefinite cache in Codex and ChatGPT Work eliminate this barrier, and caching becomes the default economics for long agent workflows: the write markup becomes a conscious calculation that is only offset by a stream of discounted reads. Manual cache optimization is commoditized at the same time—what previously required separate engineering solutions has been reduced to a single request parameter. Along with the "eternal" cache in Codex and ChatGPT Work, the cost of leaving the OpenAI platform also grows: the accumulated prompt cache remains within its infrastructure. If the pattern becomes established, TTL caching may become a standard parameter for major providers over time, and prompt design will shift to a "stable prefix plus variable tail" scheme.

Why this matters for users

For Codex and ChatGPT Work users working under their own account, the effect is immediately visible: they can calmly step away from a session for several hours and return—according to the announcement, the cache will be preserved, so returning to work no longer results in overpaying for a re-warmed context. Developers working directly with the Responses API only need to specify prompt_cache_options.ttl: "12h" in the request if system instructions and the toolset are stable throughout the workday. If requests are one-off and each has its own beginning, it is more reasonable to leave the default value "30m": paying the higher write price without repeat reads cannot be offset.

What is currently unknown / limitations

The write figures for the 12-hour cache differ between sources: the post text calls the markup "doubled" (2×), the OpenAI guide indicates 2.5×, and the example provided in the post for GPT-6.1 Sol ($2.50 at $2.00 input per 1 million tokens) arithmetically corresponds to 1.25×—the price of a 30-minute write. At least one of these statements is incorrect, and the exact value needs to be verified against the official prompt caching guide. The statement about the indefinite cache for Codex and ChatGPT Work relies on a single promotional post by Codex lead Tibo Sottiaux on X: it has not been duplicated in the changelog or guide, the storage mechanism, eviction policy, quotas, and impact on rate limits are not disclosed, and there is no independent reproduction. The phrasing "guaranteed cache hit" from the X post should be read as a marketing promise, not a documented service level guarantee.

Sources

Author

Look at AI, editorial team