Claude API
How does 90% prompt caching work on Anthropic Claude 5.5 models?
QUICK ANSWER
Anthropic prompt caching on the 5.5 model family discounts cache read tokens by 90% compared to standard input token prices. On Sonnet 5.5, cached read tokens cost $0.30 per 1M tokens (vs $3.00 base); on Opus 5.5, cached reads cost $1.50 per 1M tokens (vs $15.00 base). Cache write tokens incur a 25% surcharge for a 5-minute TTL.
5.5 Prompt Caching Economics
Prompt caching saves significant budget for iterative coding agents, video design systems, and multi-turn conversations:
- Claude Opus 5.5: Base Input: $15.00/1M • Cache Read (90% off): $1.50/1M • Output: $75.00/1M
- Claude Sonnet 5.5: Base Input: $3.00/1M • Cache Read (90% off): $0.30/1M • Output: $15.00/1M
- Claude Haiku 5.5: Base Input: $0.25/1M • Cache Read (90% off): $0.025/1M • Output: $1.25/1M
Verified against: Anthropic Developer Documentation