Claude Prompt Caching: 80% Bill Reduction in One Setting

The Math

Anthropic's prompt cache pricing for Claude Opus 4.7:

• First write (cache miss): 1.25x input price → $18.75/1M
• Subsequent reads (cache hit, 5-min TTL): 0.1x input price → $1.50/1M
• Regular input: $15/1M

Take a chat assistant with a 50,000-token system prompt that handles 100 user messages per session:

Strategy	Per-session cost	Notes
No cache	100 × 50K × $15/1M = $75.00	Full price every turn
With cache	1 × 50K × $18.75/1M + 99 × 50K × $1.50/1M = $8.36	89% reduction

For a B2B product with 1,000 active sessions per day, that is $66,640/day vs. $8,360/day — a $58K daily delta, or $1.7M per month.

TTL trade-offs

The default TTL is 5 minutes. Anthropic also offers a 1-hour TTL at 2x the write cost. Use 1-hour TTL when:

• Sessions span coffee breaks (most chat products).
• You're rate-limiting users to one message per minute.
• The system prompt rarely changes.

Use 5-minute TTL when:

• Users send messages in rapid bursts then disappear.
• The system prompt is per-tenant and tenants rotate.

Cache write granularity

Anthropic caches at the cache_control boundary you mark in the API. Place the marker after stable content (system prompt, retrieved documents) and before per-turn content (user message, conversation history). Markers are limited to 4 per request — use them on the largest stable blocks first.

Real-world hit rates

Production observations:

• Customer support bots: 80-95% hit rate (system prompt rarely changes).
• RAG with retrieval: 30-50% (retrieved chunks vary per query).
• Code agents: 60-75% (project context stable across multiple file edits).

OpenAI comparison

OpenAI's prompt caching is automatic (no cache_control needed) and bills cache reads at 0.5x — less aggressive than Anthropic's 0.1x but still significant. There's no write premium.