> ## Documentation Index
> Fetch the complete documentation index at: https://docs.youniverse.ai/llms.txt
> Use this file to discover all available pages before exploring further.
# Prompt caching
> How Aura reuses work between messages to answer faster and cost less.
Every time you send a message, the AI re-reads **everything** before it: instructions, tool definitions and the whole conversation — usually 90%+ of the input. **Prompt caching** lets the provider keep that already-processed beginning in memory, so the next message only pays full price for what is new.
## The economics
Current list prices per million tokens:
| Model | Fresh input | Cached input | Discount |
| - | - | - | - |
| Claude Haiku 5.5 | \$0.10 | \$0.01 | 90% |
| Claude Opus 5.5 | \$4.00 | \$0.20 | 95% |
| GPT-6.1 Sol | \$2.00 | \$0.10 | 95% |
| GPT-6 Luna | \$0.10 | \$0.01 | 90% |
| Gemini 3.8 Flash | \$0.75 | \$0.075 | 90% |
| Grok 4.7 | \$2.00 | \$0.50 | 75% |
Cached tokens also skip most of the compute, so the first word of the answer arrives sooner.
## Why most apps get this wrong
A cache only works if the **beginning of the prompt is byte-for-byte identical** to last time. Many chat apps break it on every message:
* They slide the history window by one message per turn, so the start changes every time.
* They insert timestamps, search results or memory near the top.
* They spread a conversation across servers, each with a cold cache.
* They mark one cache point, or mark it inside text that later changes.
## How Aura does it
Unchanging content (instructions, tools) always goes first. Anything that changes each turn — live search results, files, the time — goes at the end.
The system prompt gets a one-hour cache and a second point lands on the last settled message — never inside the newest four messages, which can still change.
Aura keeps a rolling 40-turn history but drops old messages ten at a time, so the cached beginning stays valid for many turns in a row.
Every conversation carries a stable session key, so it keeps landing on the same warm cache instead of a cold server.
Claude, Gemini and GPT cache differently (explicit breakpoints, last-breakpoint-only, or cache keys). Aura applies the right method to each.
## Measured results
* In a recent 7-day sample of logged requests, **69% of all input tokens were served from cache**.
* A repeated 17,000-token chat turn dropped from **$0.0347 to $0.0018** — about 95% cheaper.
* A repeated 26,000-token AI Builder scan dropped from **$0.21 to $0.009**.
Those savings are why Aura can offer Unlimited chat and low per-message credits.
## Keep your cache warm
* **Stay in one chat** per topic. A new chat starts cold.
* **Follow up quickly** — conversation caches last about 5 minutes; instruction caches about an hour.
* **Attach files once**, early, instead of re-uploading them.
Only briefly, in the provider's memory, scoped to your conversation, and it expires automatically. It is never used for training and is compatible with [Zero Data Retention](/learn/zero-data-retention).
No. A cache only matches an identical prompt beginning, which includes your private conversation.
This documentation is built and hosted on [Mintlify](https://mintlify.com), a developer documentation platform.