Claude Code May Be Injecting ~20K Invisible Tokens Per Request
Summary
- • Developer proxy test finds v2.1.100 adds ~20,000 undisclosed server-side tokens per request
- • Extra tokens drain paid usage limits but are invisible in the CLI context view
- • Anthropic acknowledged faster limit depletion generally but hasn't addressed the injection claim
- • Community workaround: downgrade to v2.1.98 via npx [email protected]
Details
~20,200 extra tokens per request
v2.1.98 billed 49,726 tokens; v2.1.100 billed 69,922 tokens for the same project and prompt — a delta of ~20,196 tokens
Inflation confirmed server-side
The v2.1.100 client request was smaller in bytes than v2.1.98, ruling out user-side inflation and pointing to server-side token injection
No audit trail for users
Extra tokens do not appear in the CLI /context view, are absent from Anthropic's changelogs, and cannot be audited or controlled by users
Context quality impact beyond billing
The source states (not speculates) that injected tokens enter the live context window, diluting CLAUDE.md custom instructions and degrading output quality in long sessions
Anthropic's partial acknowledgment
Anthropic acknowledged users are hitting limits faster than expected generally, but has not commented on the token injection allegation specifically; cause labeled unconfirmed
April 4 third-party tool policy change
Separately, Anthropic removed Claude subscription limit sharing for third-party tools like OpenClaw on April 4; affected users received a one-time credit equal to their monthly subscription
Community-sourced investigation into Claude Code token billing anomaly in v2.1.100
What This Means
Claude Code users on paid subscription tiers — including the $200/month Max 20x plan — may be having their limits consumed by thousands of tokens per request they never sent and cannot see, with no official explanation from Anthropic. Until Anthropic addresses the allegation, the only community-available workaround is downgrading to v2.1.98 via npx.
