Thresholds, Costs, ZDR — Prompt Caching Framework

[ ABORT TO HUD ]
SEQ. 1
SEQ. 2

Thresholds, Costs, ZDR

⚡ Prompt Caching Framework⏱ 15 min⭐ 400 BASE XP⌨ HANDS-ON LAB

Economic and Technical Limits

Writing to cache incurs a 1.25x base input rate for the default 5-minute TTL, or 2.0x base input rate for the optional 1-hour TTL ("ttl": "1h"). Subsequent Cache Reads receive massive discounts:

  • 0.025x base input (97.5% off): Claude Fable 5.1 ($0.25 / MTok)
  • 0.05x base input (95% off): Claude Opus 5.5 ($0.20 / MTok) & Claude Sonnet 5.5 ($0.10 / MTok)
  • 0.10x base input (90% off): Claude Haiku 5.5 ($0.01 / MTok ≤100K, $0.05 / MTok >100K)
  • Minimum Cacheable Length: 512 tokens across all Claude 5.x models (Fable 5.1, Opus 5.5, Sonnet 5.5, Haiku 5.5). Shorter prefixes bypass caching silently without error.

Lifecycle & Compliance

Every cache hit refreshes the 5-minute TTL timer at no extra cost. Prompt Caching is fully compatible with Zero Data Retention (ZDR)—cached KV states reside strictly in ephemeral inference memory and vaporize upon expiry.

⌨ HANDS-ON LABVerify Cache Hit Metrics & Discounts
⭐ +200 XP

Inspect cache read metrics. Fire a follow-up request matching the cached prefix on Claude Sonnet 5.5 and verify cache_read_input_tokens in usage.

1Re-send the request with the identical cached system prefix and extract usage metrics to confirm the 95% cache read discount on Sonnet 5.5.
lab-sandbox — simulated environment
INFINITY LAB SANDBOX v2.6 — simulated shell
Type the command for the current objective. Helpers: "hint", "solution", "clear".
$
OBJECTIVE 1 / 1 — type "hint" if stuck
SYNAPSE VERIFICATION
QUERY 1 // 3
If you send a 300-token prompt with cache_control to Claude Sonnet 5.5, what happens?
It caches successfully
The system ignores the cache marker and processes it normally because the prefix is below the 512-token minimum
It costs double tokens
It throws a 400 error