Claude Code Shipped Two Cache-TTL Knobs. The Long One Costs 60% More Per Write

The Claude Code 2.1.243 changelog added promptCacheTtl and subagentPromptCacheTtl — cache duration, now a setting. Before you flip either to an hour, read the price. Anthropic's prompt caching docs bill a 1-hour cache write at 2x base input against 1.25x for the 5-minute default — a 60% premium on every write. Reads stay at 0.1x either way.

That premium buys exactly one thing: survival across a gap longer than five minutes. Take 200k tokens of pinned context on Opus 5. Write it, walk away for ten minutes, come back. On the default TTL the cache is gone and you pay $1.25 twice — $2.50. On the 1-hour TTL you pay $2.00 to write and $0.10 to read it back — $2.10. One pause covers the premium. If your turns are always thirty seconds apart, you just paid 60% more for nothing.

Which is why there are two knobs, not one. Subagents that fan out, do their work, and exit never re-read their own cache — they eat the write premium and collect zero hits. The main loop, sitting idle for the eight minutes that fleet is running, is the side that pauses. Different answer per side.

Same release added a Loops breakdown to /usage with per-loop token metrics. Measure the gap between your turns before you buy the long TTL — most of the bill is already re-sent context.

See how we close it →