Claude Code burn rate: why is my usage draining so fast?
Published 29/10/2026
If your weekly meter drains faster than you expected, it is rarely the number of messages. It is usually one of three things: a long conversation re-read on every turn, a cache that had to be rebuilt, or another app on the same account.
Model a5v3, last checked against a live meter on 29/10/2026, on a Max 20x plan. More: plan usage burn data.
1. Long context costs on every turn#
Every turn re-reads the whole conversation from cache. That is cheap per token but it is paid again each turn, so a long session costs more per turn even when nothing goes wrong. Over a week of long agentic work, re-reading context was the biggest single line, bigger than output.
The fix is a smaller context, not fewer messages: keep big tool output out of the conversation, and start new work in a fresh session. (When to clear, compact or fork.)
2. The cache was rebuilt#
After a break longer than the cache lasts, a compaction or a model switch, the next turn writes the whole context again. At 800k context that one turn costs about 23× a normal one. (What a cold cache costs.)
3. The meter counts the whole account#
The terminal, the desktop and web apps, cloud sessions and a second computer all draw on the same meters. A tool that only counts one machine's requests will read low. When ours did, the missing points were always on another surface.
Your pace#
An even week is about 14.3% a day. To see whether you will reach the reset:
safe daily pace = (100% - used so far) / days until the reset
Above that pace, the week ends early. Watching that number tells you to change something while there is still week left.
Where ADE fits#
ADE records each engine call's tokens on your machine (ade telemetry), shows each engine's week against the share you set, and parks work when an engine is over it. Its habits go at causes 1 and 2: long output kept out of context, a generated handover instead of a long carried-over conversation, and small cards that finish in one window. See Day to day.