← sources

Claude caught Claude wasting tokens

source: reddit (reddit:ClaudeAI)published 2026-10-27 02:58ingested 2026-10-28 02:35

retained text

I created a voice-activated helper tool for myself. The tool works great, but it caught an agent burning tokens for no reason.

**Setup**  
  
Fable came up with a typical voice-activated CLI tool `(STT->server->Claude Code->TTS`). Opus implemented it. All local, except Claude Code acting as an agent that responds to my voice messages.  
  
**Issue**  
  
After one weird test session Fable read the transcript and came back with this:

`The model misbehaved, not the harness. After your question, Claude’s first turn ran for 72 seconds and produced about 8,000 output tokens of plain text consisting almost entirely of zero-width Unicode characters (U+200B, U+200C, U+200D), the invisible kind used for joining or breaking words. No tool call, no say, just a long invisible string, then a final “(silent)”. The harness logs plain text and speaks only \`say\`, so you heard nothing, the “waiting” indicator cleared when the turn ended [...]`

`The log line for that turn records 8,002 output tokens and an estimated cost of about $0.24 [...]`

That's a first for me. I never caught Claude producing gabage and outputting it myself. And I've never had another Claude session catch that, point fingers and put a price tag on it. Cool stuff and all, bnd I want my $0.24 back! ;) 

document fetched 2026-10-28 00:23 · canonical https://reddit.com/r/ClaudeAI/comments/1xghc4s/claude_caught_claude_wasting_tokens