Fetching from the wire…
01
02
03
04
05
06
07
08
Public story · 2026-08-15 · high
Cache reads run a tenth of input cost, but the discount expires in an hour on subscriptions and five minutes on API keys.
Why now: Anthropic published the breakdown on August 14, giving subscription users their first documented number for how long a session's cache actually lasts.
Anthropic published a breakdown of what actually drives a Claude Code session's token bill, per an August 14 post on its blog. The headline numbers: output tokens cost about five times input tokens. Decoding runs one token at a time, so it can't parallelize like reading a prompt does. Cache reads run at a tenth of the input price.
The catch is timing. On subscription plans, Anthropic says the prompt cache expires after one hour of inactivity. On API keys, it expires after five minutes. For subscription users running long agent sessions, that one-hour window sets the pace, not the model or the task list.
The post's guidance follows from the same math. Run /clear between unrelated tasks so you're not paying to reload a bloated context. Fix the model and effort level at the start of a session. Switching mid-conversation breaks the cache you already paid to build. @-mention a file directly instead of asking Claude Code to find it, which skips a Read call. Route noisy command output to a subagent, or add a quiet flag. Run /context in a fresh session to see which tool definitions are burning tokens unused. That routing advice lines up with a separate update that turned subagent forking on by default in Claude Code 2.1.232, per related coverage.
Anthropic is describing cache management as a workflow discipline, not a background optimization Claude Code handles for you. That's a real change for subscription users who leave sessions open and assume the cost sorts itself out.
Each link below shares sources, entities, or timing with this story.
Bun's Zig-to-Rust port spent 5.9 billion input tokens and 165,000 dollars, and passed the full existing test suite before merge.
The internal tool that triages Anthropic's build failures ships as reference code that mines your own incident history before it touches anything.
Output tokens cost roughly 5x input, because decode is sequential. Cache reads cost 0.1x input price. The prompt cache expires after 1 hour on subscriptions and 5 minutes on API keys. Those three numbers are the whole post, and Anthropic put them in one place for the first tim...
Clinton's July 17 guide asks four questions instead: ingestion, permissions, blast radius, observability.
The technique filters terminal output an agent reads one line of, and resolution rates held steady on 50 SWE-bench Lite tasks.
The default is 200 search calls and 200 subagent spawns per session, adjustable through an environment variable when you hit it.
MindPattern daily
One email a day at 7 AM. Sources and a take on every story. Unsubscribe anytime.