Anthropic prompt cache (cheaper repeat AI calls)
Omniscio's own AI helpers ask Anthropic to remember their long, unchanging instructions, so a repeat call is cheaper. It is on by default, each helper uses either a five-minute or a one-hour memory, and one switch in Settings turns it off for all of them.
On this page
What it is
Several helpers inside Omniscio send Anthropic the same long set of instructions on every call. Anthropic can keep such instructions in a short-lived memory, called the prompt cache. A call that finds its instructions there pays a small fraction of the usual price for them, while the first call that stores them pays slightly more.
This page is about the calls Omniscio makes itself. It is not about your Claude Code sessions: those run through the Claude program, which manages its own caching, and nothing here changes it.
What is cached
Only the part of a request that is the same on every call: the standing instructions and, for tool-using helpers, the tool definitions. Anything that differs per call (a date, a name, a document) is placed after that part and is never cached, because a cached block that changes slightly each time would be written every call and read never, which costs more than not caching.
How long it is kept
| Helper | Memory | Why |
|---|---|---|
| Plain Speak | One hour | One call at a time, the same instructions, minutes apart (cached once your rule text is long enough for the model) |
| Codex to Claude bridge | Five minutes | Many rounds in a short time; the shorter memory is enough and cheaper to store |
| Tool-using helper loops | Five minutes | Same reason |
| Coaching helpers (levels, core profile, insights, memories, compaction, summaries) | Not cached | Their fixed instructions are far under the model's minimum size, so a cache request could never take effect |
| Goal helper | Not cached | Same reason |
| Snack helper | Not cached | Its instructions are rebuilt on every call |
The one-hour memory costs more to store than the five-minute one, so it is given only to helpers that call a few minutes apart. Document text is never held for an hour.
Too small to cache
Anthropic only keeps instructions above a minimum size, and the minimum depends on the model:
| Smallest size (tokens) | Models |
|---|---|
| 512 | The newest generation: Fable 5, Mythos 5, and the 5.5 models |
| 1,024 | Opus 4.8, Opus 4.1 and 4.0, Sonnet 5.0, 4.6, 4.5 and 4.0 |
| 2,048 | Opus 4.7, Haiku 3.5 |
| 4,096 | Opus 4.6, Opus 4.5, Haiku 4.5 |
A prompt shorter than its model's minimum is sent exactly as before, with no cache request. A model that is not in the table is never cached.
Where to find it
Settings → Sessions → Cost, in the advanced spending group: Reuse long AI instructions. Settings search finds it too.
How it behaves
- On by default. The switch controls both memories at once; there is no separate setting for each.
- Off means unchanged. With the switch off, every request is identical to what was sent before this feature existed. So is a request from a helper that does not use the cache.
- Priced honestly. Every helper that uses the cache passes its cache writes and reads to the cost row, so the dearer one-hour write and the cheap read both show up in the cost you see. Nothing is hidden.
- Nothing to configure. There is no per-helper setting. A helper whose instructions fall below the minimum simply is not cached.
- Other providers. Only calls sent straight to Anthropic use this. A model reached through another provider is not cached here.
Related
- System cache cap — a different cache: the computer's file cache.
- UI usage tracking — where cost shows up.
- AI providers — the providers behind sessions.
Last verified 2026-10-30