February 20, 2026
Who stole my credits?
Excessive token usage in Claude Code
“Credit drain” panic after Claude Code update — devs clash over whether it’s real
TLDR: Some users say the latest Claude Code update suddenly burns through their credits much faster, while others report no change. The comments split between blaming a recent extension update, dismissing it as a vague complaint, and a bigger debate over pros wanting aggressive AI “help” vs hobbyists needing thrift—important if you’re hitting usage limits.
A bug report lit up the thread with a spicy claim: after updating to Claude Code 2.1.1, one user says their “credits” (usage allowance) got devoured “4x faster,” watching 10% disappear while running a simple “plan mode” in Opus. They swear nothing else changed — and note “Heiku” seemed fine — so what gives?
Cue the comments. One skeptic waved it off as “just a vague complaint,” while others piled on with it’s-not-just-you energy. Another user blamed the VS Code extension update, saying the AI helper now freezes and “the context limit fills up faster,” suggesting something changed behind the scenes. Then came the hot take that stole the show: hobbyists vs. pros. One commenter said pros want Claude to spin up extra “helper bots” and use more credits to get better results, but hobbyists on $20/month plans want it to be thrifty. Translation: “more brains” costs more — and not everyone wants to foot the bill.
Meanwhile, a couple of power users chimed in with “nothing to see here,” saying their usage hasn’t budged on Opus or Sonnet. The meme of the day? “Claude’s on a snack run” — gobbling credits like Doritos. Whether it’s a bug, a settings shift, or just heavier AI “thinking,” the community is split and loud — and that’s the drama we live for.
Key Points
- •A user observed significantly increased usage consumption after updating Claude Code to version 2.1.1.
- •The tool reportedly hits a 5-hour limit about four times faster than before, despite unchanged workflows.
- •During a plan mode run in Opus, usage rose by about 10% quickly without reading the entire database.
- •Environment details: Linux platform, gnome-terminal, version 2.1.1; a Feedback ID is provided.
- •Haiku appears less affected, with roughly 2–5% usage over nearly an hour, suggesting model-specific differences.