Claude Code 2.1.260 Will Now Tell You Why Your Cache Missed
A prompt-cache read bills at 0.1x input. Rewriting the prefix you just lost bills at 1.25x. That is a 12.5x swing on identical tokens — and until this release nothing told you which one you had paid for.
NeuroX AI · September 4, 2026
One line in the Claude Code 2.1.260 changelog is worth more than its length suggests: /cost and the status line's prompt_cache field now surface "a likely cause for prompt-cache misses (e.g. tool definitions or system prompt changed, idle past the TTL)."
Price the thing being diagnosed. Anthropic's caching docs put cache reads at 0.1x base input and a five-minute cache write at 1.25x. A miss costs 12.5x what the hit would have. On the one-hour TTL, which writes at 2x, it is 20x — the premium we priced out last week now has a matching downside.
What makes it worse is the blast radius. The cache invalidates down a strict hierarchy — tools, then system, then messages — and a change at any level takes out that level and everything after it. Edit one tool definition and the entire prefix is gone, not the tail of it. An MCP server reconnecting mid-session, a plugin registering a tool, a skill loading on demand: each one is a full rewrite of context you already paid to store.
None of that used to leave a trace. The tokens were billed as input, the session felt normal, and the only artifact was a number on an invoice weeks later. Naming the cause turns a cost line into a fixable engineering fact — which is the whole difference between a system you run and a system you can budget.