Tibo's Day 16 | 12-hour API caching and resumed Codex chats

The Day 16 post covers two related changes with different scopes: a selectable 12-hour prompt-cache option in the API, and Tibo’s statement that signed-in Codex and Work retain caches so conversations resumed hours later can hit them.

Which tasks suit a 12-hour API cache?

The @OpenAIDevs announcement concerns agent tasks that reuse instructions, tools and background throughout a day. Keeping stable material cached longer allows subsequent reuse at the lower cached-input rate.

The official documentation requires a reusable prompt-prefix match. Changing earlier instructions or tools can affect hits. A shared conversation identifier by itself does not guarantee a hit.

The current official documentation supports this option for GPT-5.6 and later in the Responses API. The request configuration fragment is:

{"prompt_cache_options": {"ttl": "12h"}}

12h specifies a minimum cache lifetime. Writes cost 2.5 times the ordinary input-token rate; this is the total write rate. Reads that hit the cache use the model’s cached-input rate. A minimum lifetime does not guarantee lower latency on every request.

Repeatedly reusing long instructions across a morning and afternoon is a stronger candidate than a single call or a task that keeps changing its prefix. Compare writes, reads and uncached input together, using the current official caching documentation for pricing. This is a mechanism-based comparison, without a CRW cost benchmark.

Do signed-in Codex and Work users need an API setting?

Tibo’s product statement concerns resuming earlier conversations in signed-in products. It does not instruct those users to configure an API cache option. Keep the API and subscription-product paths separate.

A reader asks whether subscription caching is permanent. The current basis is Tibo’s public statement about product behavior. Long-term retention, invalidation conditions and consistency across clients have not been independently verified. Cache reuse also does not enlarge a model’s context window or remove normal usage consumption.

How this extends Day 12

Day 12 describes subagent reuse, longer caching and routing improvements. Day 16 adds more specific retention options and resumed-conversation behavior. Both concern less repeated processing, without announcing a common usage-increase multiplier for all users.