Hey Anthropic, That Looks Illegal?
Tim Carambat ·2026-04-29 ·1 min read
Summary written by us from the video's transcript. The video, and everything in it, is Tim Carambat's work.
The article explains a billing bug in Anthropic’s Claude Code where certain strings in a repo trigger extra charges, why it happened, and alternatives using open‑source tools.
Takeaways
- A specific uppercase string (hermes.md) in a repo’s Git history triggered false extra‑usage charges in Claude Code’s billing system.
- Anthropic’s anti‑abuse detection relied on simple regex, leading to unintended overcharges and mixed messages about refunds.
- The incident illustrates the risks of proprietary AI services that embed opaque usage monitoring into their pricing models.
- Open‑source tools such as pi.dev and Hermes Agent can be paired with local models to sidestep such billing pitfalls.
Context and the speaker’s stance on AI tools
Tim Carambat introduces his upcoming talk at AMD AI Dev Day, emphasizing that AI should be treated as a tool—not a replacement for human work. He argues that the narrative of AI replacing people is overstated and that developers benefit more from using AI as an assistive instrument.
Claude Code’s recent billing controversy
Carambat recounts how Anthropic briefly removed Claude Code from its base‑tier plans in an AB test, then reinstated it after backlash. He notes a subsequent issue where users on the highest 20X plan received “out of extra usage” notices despite being well below their allotted quota.
The specific bug: string detection in Git history
A Reddit user discovered that having the exact uppercase string **hermes.md** (or similar competing harness identifiers) anywhere in a repository’s Git history caused Claude Code’s billing system to flag extra usage. The bug was reported on GitHub, and Anthropic admitted it stemmed from an over‑active anti‑abuse regex that mistakenly treated the string as a trigger for overage charges.
Anthropic’s response and customer impact
Anthropic initially said they would refund affected users and provide credits, but internal comments suggested they intended to keep the money. The head of growth labeled it an AB test, while the creator of Claude Code called it an “overactive anti‑abuse system” that has since been fixed.
Why such heuristics are problematic
Carambat critiques the reliance on simple regex checks of commit histories as a billing safeguard, noting that Anthropic also uses similar regex for sentiment analysis. He argues this approach is unsustainable and borders on fraudulent practice when it leads to unwarranted charges.
Alternatives: open‑source harnesses with local models
The video highlights open‑source options like **pi.dev**, **Hermes Agent**, and other community tools that can run against local models, avoiding the opaque billing mechanisms of Claude Code. Carambat encourages developers to adopt these alternatives to regain control and avoid unexpected fees.