Skip to content
Hire an AI employee
AI EmployeeSuper-AgentsAgent-to-AgentBenchmarksPricingBlog
Insights 03 Cost, ROI & pricing

OpenAI 12-Hour Prompt Caching: When It Pays Off

Nitish Garg · Founder & CEO, CellCog · 20 October 2026 · 5 min read · 1,048 words
At a glanceQuick answers
What changed?
You can now ask OpenAI to keep a cached prompt prefix for at least 12 hours instead of 30 minutes, by setting a ttl of 12h in the Responses API.
When should I use it?
When the same instructions, tools or reference material get reused with gaps longer than 30 minutes. For back-to-back calls, the cheaper 30-minute cache stays warm on its own.
What is the catch?
Writes cost twice as much as a 30-minute write, upgrading an existing entry charges a fresh write, and it only works in the Responses API on GPT-5.6 and later.
Data illustration on a near-white ground: a day timeline with twelve hourly ticks under one long lit cache bar, a single large write block at the start and thin read slivers after it, under the large figure 12 h
Fig 0One write, a day of cheap reads. Made by CellCog's image agent, running GPT Image 2.5.

OpenAI now lets developers keep a cached prompt for 12 hours instead of 30 minutes, at a higher price to write it and the same cheap price to read it. The API changelog lists it on October 20, 2026: “You can now choose 12-hour prompt caching in the API.” OpenAI Developers announced it on X at 17:56 UTC.

The pitch is aimed at agents that run all day: “Persistent agents, like dots, reuse context throughout the day.” This page reads OpenAI’s prompt caching guide and pricing page directly and works through what the change costs.

On this page · 6 sectionsOpen
  1. What changed
  2. When it pays off
  3. The fine print
  4. Why OpenAI shipped it now
  5. What we are watching for
  6. Sources
Key points5 · 5 min full read
  1. A clock face over a stack of cards.
    OpenAI added a 12-hour prompt cache to the Responses API on October 20, 2026, for GPT-5.6 and later models.
  2. Two price tags of different sizes.
    A 12-hour cache write costs 2.5 times the normal input price, against 1.25 times for the default 30-minute cache; reads cost the same either way.
  3. Three ascending steps with a checkmark.
    Against no caching, the 12-hour cache pays for itself from the third use of the same prefix within its lifetime.
  4. A falling bar chart.
    In our worked example, an agent calling once an hour with a 100,000-token prefix on GPT-6.1 Sol pays $0.61 a day instead of $2.40, about 75% less.
  5. A small robot beside a sun and moon.
    OpenAI pitches it at persistent agents like its own dots, which reuse the same instructions and tools all day.

§ 01What changed

A prompt cache stores the start of a request (system instructions, tool definitions, reference documents) so later requests that begin the same way are billed at a much lower cached rate. Until now the cache lived for a minimum of 30 minutes. An agent that wakes once an hour found it gone and paid to write it again.

The new option keeps the entry for at least 12 hours. The changelog states the price: “12-hour cache writes cost 2.5x the input token price.” The guide adds that these “are total write rates, not surcharges on top of ordinary input.” Reads cost the same for both lifetimes.

Model Input Cached read 30-minute write 12-hour write
GPT-6.1 Sol $2.00 $0.10 $2.50 $5.00
GPT-6 Astra $10.00 $1.00 $12.50 $25.00
GPT-6 Luna $0.10 $0.01 $0.125 $0.25
Table 1OpenAI cache prices per 1M tokens, short context (OpenAI pricing page, October 20, 2026)

Older models such as GPT-5.5 have no write prices on the page; the 12-hour option is for GPT-5.6 and later, in the Responses API.

§ 02When it pays off

The arithmetic is simple. A 12-hour write costs 2.5 times a normal input, and every later read costs 0.1 times. Against sending the same prefix uncached each time, the 12-hour cache wins from the third use: two uncached sends cost 2.0, while a write plus one read costs 2.6; at three it is 3.0 against 2.7.

Against the 30-minute cache, the question is the gap between calls. If calls come every few minutes, the 30-minute entry stays warm on its own and the cheaper write wins. If calls come more than 30 minutes apart, the 30-minute entry can expire between them and every call pays a fresh write.

Setup GPT-6.1 Sol GPT-6 Astra
No caching $2.40 $12.00
30-minute cache, rewritten every call $3.00 $15.00
12-hour cache, one write then 11 reads $0.61 $3.60
Table 2Worked example: a 100,000-token prefix, one call an hour for 12 hours (our arithmetic from OpenAI’s prices)
Worked example: 12 hourly calls, 100k-token prefix, GPT-6.1 Sol ($)Bar chart: no caching 2.40, 30-minute cache rewritten every call 3.00, 12-hour cache 0.61, with the 12-hour cache highlightedNo caching2.4030-minute, rewritten3.0012-hour cache0.61Worked example: 12 hourly calls, 100k-token prefix, GPT-6.1 Sol ($)Bar chart: no caching 2.40, 30-minute cache rewritten every call 3.00, 12-hour cache 0.61, with the 12-hour cache highlightedNo caching2.4030-minute, rewritten3.0012-hour cache0.61
Fig 1Worked example: 12 hourly calls, 100k-token prefix, GPT-6.1 Sol ($)

On GPT-6.1 Sol the 12-hour cache cuts that agent’s prefix cost by about 75% against no caching, and by about 70% on GPT-6 Astra. The middle row is the worst case: OpenAI notes it “may retain entries longer than the requested minimum,” so a 30-minute entry sometimes survives an hour.

Animation: a day timeline with twelve hourly calls; the top lane, 30-minute cache, adds an amber write at every call while its cache bubble expires between them and its counter climbs to $3.00; the bottom lane, 12-hour cache, makes one teal write then thin reads under a bubble that stays lit, stopping at $0.61, beside a grey no cache line at $2.40
Fig 2An hourly agent: twelve fresh writes with a 30-minute cache, one write and eleven cheap reads with a 12-hour cache

§ 03The fine print

  • Upgrading costs a full write. “Upgrading valid 30-minute coverage to 12 hours charges a new 12-hour write for those tokens, without credit for the earlier write.”
  • You can mix lifetimes. Explicit breakpoints let you put stable instructions on a 12-hour entry and the changing conversation on a 30-minute one, and each request can create up to four cache writes.
  • Responses API only. The older prompt_cache_retention setting for earlier models is unchanged.
  • Retention. The guide says the 12-hour setting is a minimum lifetime, not a deletion deadline, and does not change OpenAI’s maximum data-retention policy.

§ 04Why OpenAI shipped it now

Persistent agents changed the shape of API traffic. A chatbot sends bursts of requests in one sitting; an agent like OpenAI’s own dots wakes up on a schedule, checks something, and goes quiet for an hour. OpenAI’s Nikunj Handa put it plainly: “Some persistent agents (similar to dots) building on our API have seen big cost savings via this feature.” The 12-hour cache is the billing model catching up with agents that work a full day.

§ 05What we are watching for

  • Whether Anthropic and Google answer with longer cache lifetimes of their own.
  • Whether OpenAI brings the 12-hour option to the Chat Completions API.
  • Real bills from teams running scheduled agents, which will show how often the 30-minute cache was already surviving past its minimum.

§ 06Sources

Frequently asked5 questions

Q1When did OpenAI launch 12-hour prompt caching?

October 20, 2026. It appears in the API changelog for that date, and OpenAI Developers announced it on X at 17:56 UTC.

Q2How do I turn it on?

In the Responses API, set prompt_cache_options.ttl to 12h for the whole request, or set a ttl of 12h on an individual explicit cache breakpoint. Leaving both out keeps the 30-minute default.

Q3Does the 12-hour cache make reads cheaper?

No. Reads cost the same for both lifetimes: 0.1 times the input price on most models and 0.05 times on GPT-6.1 Sol. You pay more once to keep the entry longer.

Q4Does 12-hour caching change data retention?

OpenAI’s guide says the 12-hour setting is a minimum, not a deletion deadline, and does not change its maximum data-retention policy.

Q5Does CellCog use OpenAI's prompt caching?

No. CellCog’s agent tiers run on Claude Opus 5.5. We follow OpenAI’s pricing because it sets what businesses expect a persistent agent to cost.

Published 20 October 2026 All Cost, ROI & pricing →