Claude Moved My Cheese

Claude and I had a great system in place for me to quickly and easily assess how many tokens he had left before auto-compacting. This made it easy for me to see if he had enough tokens remaining to take on a new task. Or if he was in the middle of something, I could decide if it was safe to continue past compaction, or if it would be better to start a new session instead. At the end of each task, or when he felt he’d completed a big chunk and wanted to check in with me, he’d display something like this:

Claude D – Token Usage: 87K / 120K (72% used)

(The “120k” was our safety margin—he auto-compacted at around 130k.)

Since I work with up to ten instances of Claude Code at a time on different tasks, being able to glance at a terminal window and see token usage reports made it super easy to see where a session was at. At the time, Claude explained that he got regular system warnings about his usage, and based on those and whatever work he’d done since the last warning, was able to make pretty accurate estimates. 

Alas, this no longer works.

Recently I started to notice all the Claudes were way off on their token usage reports. Yesterday Claude H reported that he’d used ~25k tokens, but when I ran the /context command it reported he’d spent 61.7k tokens on messages alone. I asked a Claude session about this, and he said he was estimating his token usage (and clearly not estimating very well), and when I asked about system warnings he said he doesn’t get any. The default model is now Opus 4.5 (previously Sonnet), so I’m guessing either Opus works differently, or else Anthropic has changed their approach. Whatever the reason, one thing is clear:

Our super awesome token usage tracking system is defunct. 🙁

This isn’t terrible. I can still run /context and get a detailed report. But it’s slower—I have to type the command and wait for the answer. It was much nicer—and faster—to get this information proactively from the Claudes.

I’ll make do, but I’ll also occasionally ask Claude if he’s able to accurately estimate his token usage on his own again. Because one thing is for certain with AI: things are not going to stay the same.