Service eligibility and regional restrictions
PuppyIP serves only compliant overseas businesses and their authorized personnel. Proxy services are not available in mainland China. The service may only be used for lawful business activities outside mainland China. Use of this service within mainland China is prohibited.
Hosting a proxy IP or server overseas does not change these restrictions. The service must not be provided to end users in mainland China through relaying, forwarding, sharing or resale. Before use, read the Terms of Service.
Key Takeaways
- OpenAI’s October 20 API changelog lists 12-hour prompt caching as a Responses API update.
- The current guide says supported Responses API requests on GPT-5.6 and later can set ttl; 12-hour cache writes cost 2.5 times the uncached input-token rate.
- The 2.5x figure is the rate for eligible cache writes, not an extra surcharge added on top of the ordinary input price; actual results depend on cached reads and writes.
- Thirty minutes and 12 hours are minimum cache lifetimes, not hit guarantees or exact deletion times; compare usage fields with current model prices.
What changed
On October 20, 2026, OpenAI’s API changelog announced 12-hour prompt caching under the Responses API. OpenAI describes it as a way to keep shared instructions, tools, and context cached between requests that may be hours apart, so reused input can receive the lower cached-input price. The same changelog entry says 12-hour cache writes cost 2.5 times the input-token price.
This update concerns an API prompt-caching setting. It does not change every API model or product interface; confirm support for the exact model and request type before using the option.
How 12-hour cache pricing works
The current prompt-caching guide scopes the configurable prompt_cache_options.ttl to the Responses API on supported GPT-5.6 and later models. The default 30-minute option has a cache-write rate of 1.25 times the uncached input-token rate; the 12-hour option has a cache-write rate of 2.5 times that rate. Confirm support for the specific model and interface before applying these rules elsewhere.
The multiplier describes the rate for tokens written to the cache. It does not mean charging the ordinary input rate and then adding another 2.5x on top. Nor is it a multiplier for the whole request: calculate uncached input, cached reads, cache writes, output, and tool charges separately. Cache duration and this rate alone cannot show that a workload will save money.
Twelve hours does not guarantee a hit or deletion at expiry
OpenAI defines the 30-minute and 12-hour values as minimum cache lifetimes, and says cache behavior, retention settings, and prefix reuse vary by model. Cache entries are not stored indefinitely. A 12-hour setting is a minimum retention period for the selected write; it does not promise that a request will read from cache, and it is not a command to delete the entry exactly 12 hours later.
A stable shared prefix can help with reuse, but the guide says that maintaining a session does not itself guarantee a cache hit. Record the selected ttl for each request, inspect response usage such as cached_tokens and cache_write_tokens, and compare those with latency and model prices. These general usage fields alone do not prove the actual hit rate, performance improvement, or savings from a 12-hour setting.
How developers can check the cost
First confirm that the request uses a model and the Responses API that support ttl. Consider 12 hours only for stable prefixes that will be reused across longer gaps. Keep the model, request type, ttl, request timing, and existing usage data so that changing several variables at once does not obscure the cause of a cost change.
Compare a representative set of real usage: separate uncached input, cached reads, and cache writes; inspect the write-token breakdown for 30-minute and 12-hour lifetimes; then calculate the total using the model’s current prices. If a prefix changes often, requests rarely reuse it, or cache writes increase, a longer lifetime may not reduce total cost. This article did not make API calls, check account eligibility, or measure a cache-hit rate. Search demand also remains unverified because no actual query rows were available.
Sources
Frequently Asked Questions
Does 12-hour prompt caching apply to every OpenAI model?
Do not assume so. The current guide lists the 30-minute and 12-hour prompt_cache_options.ttl settings for supported Responses API requests on GPT-5.6 and later. Check the exact model and interface first.
Does the 2.5x rate multiply the price of the whole API request?
No. It is the rate for tokens written to the 12-hour cache relative to uncached input tokens. Calculate uncached input, cached reads, output, and other charges separately using actual usage and applicable prices.
Does setting 12 hours guarantee a cache hit?
No. Twelve hours is the minimum cache lifetime for the selected setting, not a hit guarantee. Check cached-read and cache-write tokens in response usage and evaluate actual prefix reuse.
Will longer caching always lower total cost or latency?
No guarantee. The 12-hour option has a higher cache-write rate. Whether it is worthwhile depends on cached reads, writes, prefix reuse, and total request volume. Compare real usage with current model prices.