Tag: prompt-caching

Claude Code 2.1.260 adds a likely-cause diagnosis to the Prompt cache (main) line in /usage and to the prompt_cache status line object. Instead of just counting misses, it tells you whether the tool set changed, the system prompt changed, or the TTL expired.

Conversation history can live in your process, in your database, in the provider, or on the machine that ran the agent. Where you put it does not change your token bill, it decides your retention exposure, and it decides whether you can ever change providers. The numbers, the retention windows, and a storage shape that survives all three.

The 1M-token context window is now the default on Claude Opus 5, Opus 4.8/4.7/4.6, Sonnet 5, and Sonnet 4.6, with no beta header and no long-context premium. Here is when deleting the vector store actually pays off, the 10x cache-read rule that decides it, the ~30% tokenizer inflation that breaks your sizing estimate, and the seven-step migration with a verification line on each one.

The cache read and write multipliers are identical on both models, so the break-even point is the same. What differs is the minimum cacheable prefix (1,024 vs 4,096 tokens), the per-token dollar savings, and a new Opus 4.7 tokenizer that counts up to 35% more tokens. With claude-sonnet-4-6 and claude-opus-4-7 pricing math.

A 429 rate_limit_error on claude-sonnet-4-6 in a long agent loop is almost always ITPM, not RPM. Read retry-after, cache the system prompt, and gate on anthropic-ratelimit-input-tokens-remaining. Step-by-step fix with code.

A 2026 walkthrough for wiring RAG into Claude Code 2.1.x: when agentic grep stops scaling, how to attach a hybrid BM25 + dense vector MCP server, how to wrap a retrieval CLI in a custom skill, and how Anthropic's contextual embeddings technique pushes recall above 92%. Anchored to claude-sonnet-4-6, claude-opus-4-7, and Claude Context 0.x.

Place rolling cache_control breakpoints across messages, respect the 20-block lookback, and refresh the 5-minute TTL automatically so a 50-turn agent loop pays the prefix once, not fifty times. Verified against anthropic 0.42 (Python) and @anthropic-ai/sdk 0.30 (Node) in May 2026.

Add prompt caching to a Python or TypeScript Anthropic SDK app, place cache_control breakpoints correctly, and read cache_read_input_tokens and cache_creation_input_tokens to compute a real hit rate. With pricing math for Claude Sonnet 4.6 and Opus 4.7.