PROMPT CACHING

Caching a shared prompt cut our worked example by about 78-80% on four models; batch takes 50% off at most providers but 40% at Moonshot and 20% at xAI. Only Anthropic, Fireworks and xAI say the two stack. The math, the assumptions and the quotes.

Prompt caching across OpenAI, Anthropic, Gemini, DeepSeek, Mistral, xAI and Qwen: prices, TTLs, setup code. A reused 100K prefix costs 71-93% less.

Claude API inline tools beta: define, reshape or upgrade a tool mid-conversation with one system message and keep the prompt cache. Code included.

Claude Code 2.1.299 compacts sessions over 200K tokens after about 54 idle minutes. How the cache math works, the setting to turn it off, and when to keep it.

Claude Code now reports prompt cache hit ratio and miss causes in /usage. How to read it, script a status line, and set 5m or 1h TTLs.

Claude Fable 5.1 keeps $10/$50 pricing but cuts cache reads to $0.25/MTok. The three breaking API changes, new betas, and when Opus 5.5 is enough.

Context engineering for agents, explained: caching, compaction, tool-result clearing, retrieval and subagents, with the exact Anthropic and OpenAI API knobs.

Fable 5 prompt caching economics: cache-write vs cache-read pricing, 5-minute vs 1-hour TTL break-even math, and worked agent-loop examples.

Cut Claude API spend by up to 90% with prompt caching. Real numbers, TypeScript SDK code, and the gotchas Anthropic's docs gloss over.

Automatic reuse of cached context for substantial cost reduction.