Prompt caching
LiveStable prompt prefixes are cached. Faster responses, lower bill, passed to you.
- ✓Stable-prefix caching against Anthropic's prompt cache
- ✓Cache-read and cache-creation tokens broken out in your usage figures
- ✓Cache reads counted at a tenth against your monthly token allowance
- ✓The cache lifetime is Anthropic's ephemeral default, not a LexVio setting
What it is.
We cache the stable prefix of our prompts — system instructions, jurisdiction context, firm style — against Anthropic's prompt cache, sending the stable half marked cacheable and the variable half unmarked. A cache read is billed by Anthropic at a fraction of a fresh input token, and it counts against your monthly allowance at a tenth.
Your usage breakdown shows cache-read and cache-creation tokens separately from ordinary input and output, so you can see the caching working. Be aware of one limit in the cost figure: the cost estimate is computed from input and output tokens and does not currently price cache reads or cache writes, so the saving shows up honestly in the token counts and approximately in the money.
Three steps.
End to end.
Cache miss — full cost, full latency.
Cache hit on the stable prefix — ~10% cost, ~half latency.
Cached vs uncached tokens broken out per request, so you can verify the savings.
What you get.
- ✓Stable-prefix caching against Anthropic's prompt cache
- ✓Cache-read and cache-creation tokens broken out in your usage figures
- ✓Cache reads counted at a tenth against your monthly token allowance
- ✓The cache lifetime is Anthropic's ephemeral default, not a LexVio setting
Quick answers.
No — only the stable prefix (system instructions, jurisdiction context, firm style). Your document content is in the variable half and is not marked cacheable.
Not precisely. The cost estimate is derived from input and output tokens and does not price cache reads or writes. The token breakdown is the accurate view of what caching saved you.
More in AI Controls & Guardrails.
Every citation comes back with a verdict — verified, ambiguous or unresolved — and a match on title alone is never verified.
Each legal claim we hold about a jurisdiction is checked against that government's own legislation service — and where it cannot be, the answer is "unverified", not silence.
We do not train models on customer content. There is no opt-in, because there is no training programme.
User-pasted text is sanitised before reaching the LLM.