Back
GitHub / AnthropicSecurityGitHub / Anthropic2026-04-29

The String 'HERMES.md' in Your Git History Can Silently Drain Your Anthropic Budget

A bug in Claude Code routes requests to pay-as-you-go 'extra usage' billing instead of your Max plan quota when the string HERMES.md appears in recent git commit messages — silently charging some users over $200 without warning.

Original source

On April 25, 2026, a Claude Code user reported that their Max plan subscription had been silently charging $200+ in "extra usage" fees while their included plan quota sat barely touched. The root cause: the string **HERMES.md** in a recent git commit message was triggering Claude Code's billing routing logic to classify the session as third-party harness usage, bypassing the flat-rate Max subscription entirely.

The trigger isn't a file named HERMES.md on disk — it's the string appearing anywhere in recent commit messages, which Claude Code includes in its system prompt context. Server-side logic appears to pattern-match this as a signal of the Hermes Agent, a popular third-party harness built by NousResearch. Starting April 4, 2026, Anthropic began rerouting flat-rate Max plan sessions to extra-usage billing when it detected third-party harness usage, citing unsustainable consumption patterns.

The problem is the classifier's blast radius. Any project that happens to reference HERMES.md — documentation files, dependency changelogs, commit messages discussing the tool — could be silently reclassified as third-party usage. Anthropic Head of Claude Code acknowledged the issue as a false positive and committed to fixes, but the damage for affected users was already done: billing errors were reported as generic "out of extra usage" messages with no indication of why the rerouting occurred.

Theodore Browne (Theo from t3.gg) summarized the community reaction: *"It is genuinely insane that Anthropic will bill you differently if you mention certain words in your prompt or have certain files in your codebase."* The incident reignited debate about the April 4 policy itself, which many developers felt was introduced without adequate communication or grace period.

Panel Takes

The Builder

The Builder

Developer Perspective

This is a billing system bug that hits developers who are doing exactly the right thing — maintaining clear commit history. The fact that a string in your git log can change what pricing tier you're on is an architectural smell that should have been caught in QA. Anthropic needs deterministic billing rules with clear documentation, not classifier-based routing.

The Skeptic

The Skeptic

Reality Check

This is a symptom of a deeper policy problem: Anthropic built a usage restriction system on top of a content-aware classifier without auditing the false-positive blast radius. Any keyword-based billing trigger is going to have this failure mode. The fix isn't just patching this bug — it's rethinking whether opaque classifier-driven billing routing should exist at all.

The Futurist

The Futurist

Big Picture

The underlying tension here is real: flat-rate AI subscriptions are economically unsustainable for heavy agentic use, and Anthropic's attempt to draw a line between 'normal' and 'harness' usage will only get messier as the lines blur. This incident is a preview of the coming subscription model reckoning across all major AI providers.

Bookmarks

Loading bookmarks...

No bookmarks yet

Bookmark tools to save them for later