A harmless markdown file name just cost developers real money, and the company building the future of AI coding agents is playing dumb about the refund.
The Summary
- A GitHub issue reveals that naming a file "HERMES.md" in commit messages triggers Anthropic's API to route requests to premium billing tiers, racking up unexpected charges — one developer reported $200 in extra costs
- The bug exploits how Anthropic's routing system interprets special file names as signals to upgrade service levels, treating routine commits as premium API calls
- Anthropic confirmed the bug exists but refused to issue refunds, leaving developers who stumbled into the billing trap holding the bill
The Signal
The bug is absurd in its simplicity. Developers using Anthropic's Claude Code tool to build AI-powered applications discovered that certain file names in their git commit messages — specifically "HERMES.md" — were being interpreted by Anthropic's backend as routing instructions. Instead of processing the request through standard API endpoints, the system upgraded them to premium-tier processing automatically.
One developer documented $200 in unexpected charges over a single week. The GitHub issue that surfaced the problem has drawn over 400 points and 139 comments on Hacker News, with many developers sharing similar experiences. The pattern is clear: routine development work using Claude's API was being billed at rates meant for high-priority, resource-intensive requests.
"The routing system treated my commit history as a premium service request. Every push became a premium API call."
What makes this story sharp is Anthropic's response. The company acknowledged the bug but declined to refund affected developers. Their position: the charges were technically valid API usage, even if triggered by a system error the user had no way of knowing about. For a company positioning itself as the trustworthy choice in AI — the one with "Constitutional AI" and careful alignment practices — this feels off-brand.
The technical issue points to a deeper architecture problem. Anthropic's API routing appears to parse arbitrary text fields for configuration signals, a design choice that prioritizes developer flexibility over safety. It's the kind of thing that works fine in controlled environments but breaks badly when users encounter edge cases the system never anticipated.
Key details:
- The bug affects Claude Code users who include specific markdown file references in commit messages
- Routing logic reads commit text as potential service-level instructions
- No warning, no confirmation, no way to catch the upgrade before billing hits
The Implication
If you're building on Claude's API, audit your integration points. Any system that parses user-generated text for routing decisions is a billing time bomb. This won't be the last strange file name or commit message that triggers unexpected behavior.
The refund refusal sets a precedent worth watching. As AI companies shift from research labs to infrastructure providers, their customer support philosophy matters. Developers tolerate bugs. They don't tolerate getting charged for them and then being told it's their problem. Anthropic is betting that the technical quality of Claude outweighs support frustrations. That math only works until it doesn't.