← Guides · November 5, 2026
Prompt Caching Explained: Cut Your LLM Input Costs by Up to 90%
If your application sends the same long instructions with every request, you are probably paying full price for text the model has already seen. Prompt caching fixes that.
What gets cached
When a request starts with exactly the same tokens as a recent request, the provider can reuse the work it already did on that prefix. Those tokens are billed at a much lower “cache read” rate. Typical cacheable content:
- the system prompt and style rules,
- tool and function definitions,
- a long reference document you ask several questions about,
- the earlier turns of a conversation.
How much cheaper is it?
Cached input prices at list rates (November 2026):
| Model | Fresh input / 1M | Cached input / 1M | Discount |
|---|---|---|---|
| Claude Sonnet 5.5 | $2.00 | $0.10 | 95% |
| GPT-6.1 Sol | $2.00 | $0.10 | 95% |
| Gemini 4 Argon | $2.00 | $0.10 | 95% |
| Claude Haiku 5.5 | $0.10 | $0.01 | 90% |
Output tokens are never cached. Caching only lowers the input part of the bill.
The one rule: stable prefix first
Caches match from the start of the prompt. Anything that changes (the user’s question, a timestamp, a random ID) must come after everything that stays the same. A common mistake is putting today’s date at the top of the system prompt, which breaks the cache on every request.
A cache-friendly order:
- System prompt and rules
- Tool definitions
- Static documents
- Conversation history
- The new user message
When caching does not help
- Very short prompts: below the provider’s minimum cacheable length, nothing is cached.
- Low traffic: caches expire after a few minutes without use, so a request every hour will usually miss.
- Output-heavy workloads: if most of your cost is output, caching barely moves the total.
Estimate your savings
In the cost calculator, set Cached input to the share of each prompt that is stable (often 30–70% for chatbots and agents) and compare the monthly totals with caching at 0%.