tokevenWhy caching matters
On most Anthropic, OpenAI and Google models, cache-read tokens are billed at 0.10x the input rate; xAI's Grok models discount less, and some providers charge a premium to write the cache in the first place. The discount is per model, so check the model page for yours. Structure prompts to maximize cache hits and the repeated part of your input gets much cheaper.
How to improve cache rates
1. Put stable content first
System prompts, instructions, and context that doesn't change between calls should go at the beginning of your prompt. The model caches from the start of the prompt.
2. Reuse conversation prefixes
If multiple calls share the same context (e.g., a code review session), structure them so the shared prefix is identical.
3. Batch similar requests
Group similar tasks together in the same session. The shared context gets cached once and reused.
Tracking your progress
The Tokens tab on Spend shows cache-read and cache-write tokens separately from fresh input, and the Caching tab on Save shows cacheable patterns in your usage. Aim for 30%+ of input served from cache for meaningful savings.