Claude Code v2.1.100 Token Inflation is a reported server-side regression in which 📝Claude Code v2.1.100 bills roughly 20,000 additional tokens per request compared to v2.1.98, despite sending fewer bytes on the wire.
The regression was surfaced by an independent researcher who ran an HTTP proxy against four consecutive Claude Code versions to capture full API traffic. On an identical project, prompt, and account, v2.1.98 billed 49,726 tokens while v2.1.100 billed 69,922 tokens — a ~20K delta attributable entirely to server-side injection, since the outbound payload from v2.1.100 was smaller. The added tokens are invisible to the user: they do not appear in /context, are not reflected in the client-side transcript, and still enter the model's working context window. Downstream effects include higher billing, dilution of user-authored CLAUDE.md instructions, and faster quality degradation in long sessions. The fix is to downgrade to v2.1.98 until Anthropic addresses the regression.
Key Facts
- Category: CLI regression / billing issue
- Affected product: 📝Claude Code CLI
- First affected version: v2.1.100
- Last known-good version: v2.1.98
- Measured delta: ~20,196 tokens per request (49,726 → 69,922 on identical input)
- Origin: server-side; client payload is smaller in the affected version
- Visibility to user: none — absent from
/contextand transcript
How It Works
The inflation is injected after the client request reaches Anthropic's infrastructure, which is why proxy captures show v2.1.100 sending fewer bytes than v2.1.98 while being billed for more tokens. Because the injection happens upstream of the model, the added tokens occupy space in the active context window alongside the user's system prompt and CLAUDE.md contents.
- Billing impact — every request on v2.1.100 incurs ~20K additional billed tokens, independent of prompt size
- Context dilution — CLAUDE.md instructions and user memory compete with ~20K tokens of unknown content, reducing adherence to user rules
- Audit opacity —
/contextshows only client-side inputs, so the inflation cannot be inspected, measured, or attributed from within the CLI - Session decay — long sessions hit auto-compact and quality thresholds sooner because the hidden payload consumes budget each turn
Why It Matters
For users on metered billing, the regression is a silent ~40% cost increase on a fixed workload. For users relying on CLAUDE.md to enforce workflow rules, it is an unaccounted-for context contaminant: when Claude ignores a rule, the user cannot tell whether the rule was overridden by invisible upstream content or by something they authored. Both effects undermine the sovereignty premise of a local CLI — that the user controls what the model sees.
FAQ
What is the Claude Code v2.1.100 token inflation?
A server-side regression in Claude Code v2.1.100 that adds approximately 20,000 billed tokens per request compared to v2.1.98, despite the client sending fewer bytes. The extra tokens enter the model's context window but are not shown in /context.
How is the inflation detected?
It is only detectable by capturing full API traffic through an HTTP proxy and comparing billed token counts across versions on identical inputs. The Claude Code CLI itself does not expose the injected content.
How do I fix it?
Downgrade to v2.1.98 by installing the pinned npm package: npm i -g @anthropic-ai/[email protected]. On systems using the native installer, repoint ~/.local/bin/claude at the npm-installed binary and set DISABLE_AUTOUPDATER=1 to prevent the updater from reinstalling a newer version.
Does the inflation affect CLAUDE.md adherence?
Yes. The ~20K injected tokens share the same context window as CLAUDE.md, reducing the relative weight of user-authored instructions and making rule-following less reliable in long sessions.
Is the fix reversible?
Yes. Once Anthropic resolves the regression, re-enable auto-updates by removing DISABLE_AUTOUPDATER from the environment and ~/.claude/settings.json, then allow the native installer or npm to move forward.
Related
- 📝Claude Code — the CLI affected by the regression
- 📝Claude Code Hooks and Skills — the configuration surface where
DISABLE_AUTOUPDATERandminimumVersionare pinned - 📝MythOS Claude Code Skill — the MCP integration that runs inside affected and pinned sessions alike