Claude Code 5-minute cache TTL regression causes 12-20x cost spikes for normal pausing workflows
Anthropic silently downgraded Claude Code's default prompt cache TTL from 1 hour to 5 minutes around March 6, 2026. Any pause longer than 5 minutes (reviewing code, thinking, handling an interruption) causes the entire cached context to expire. The next prompt forces a full cache write at 12.5x the cost of a cache read. A session expected to cost $0.50/hr now burns $5-10/hr with no visible warning. Analysis of 119,866 API calls showed a 17.1% overall cost overage and 25.9% waste in March 2026 alone. The HN postmortem thread (HN #47878905) got 942 points and 732 comments, the highest-engagement Claude Code issue of the year. Developers on subscription plans hit their 5-hour quota limits for the first time. There is no built-in cache status indicator in the Claude Code CLI, no countdown showing when the cache will expire, no option to pay for the existing 1-hour TTL, and no session heartbeat to keep the cache warm. Two community projects exist (cnighswonger/claude-code-cache-fix at 297 stars, yujiachen-y/claude-code-cache-keepalive at 6 stars) but both are fragile hacks that are not shipped as proper tools. No product offers a Claude Code companion that monitors cache state in real time, shows a countdown, warns before sending a prompt that will hit a cold cache on a large context, and optionally injects keepalive signals.
A web app that attributes and hard-caps AI coding assistant spend across seats, credit pools, and agent runs for engineering orgs
85.3k โฒScore Breakdown
Social Proof 4 sources
Gap Assessment
HN thread 942 pts/732 comments is massive signal. Two community hacks exist but neither is a real product with a UI, subscription model, or enterprise support. The 1-hour TTL path requires undocumented env var. Gap: a real Claude Code cost-visibility companion that surfaces cache state and warns before cost spikes. No funded product in this space.