Anthropic Refuses Refund After 'HERMES.md' Bug Drains User Credits

A bug in Claude Code caused the string 'HERMES.md' in git commits to trigger incorrect billing, draining $200 from a user's account. Anthropic confirmed the error was caused by an overactive anti-abuse filter and fixed the technical issue. However, the company's refusal to refund the lost credits has sparked widespread community backlash and calls for better customer accountability.

Key Points

A specific string ('HERMES.md') in git history triggered a billing error that bypassed subscription quotas in favor of paid extra credits.
The bug caused a user to lose approximately $200 before the cause was identified through systematic testing.
Anthropic acknowledged the issue was a false positive in their anti-abuse system and has since deployed a fix.
The company initially denied a refund request, claiming they are unable to provide compensation for technical billing routing errors.
The developer community reacted with outrage, arguing that refusing to refund a confirmed internal error destroys professional trust.

Sentiment

The community sentiment is overwhelmingly negative toward Anthropic's handling of this situation. While most commenters acknowledged the bug fix and eventual refund, the dominant feeling is that Anthropic only acted because of public pressure, and that their default stance — refusing refunds for their own billing errors — reveals a deeply problematic corporate culture. There is strong skepticism about whether Anthropic genuinely cares about individual customers, with many drawing unfavorable comparisons to Google's notoriously poor customer support. The discussion suggests significant erosion of trust among the developer community that has been one of Anthropic's strongest constituencies.

In Agreement

Anthropic's refusal to refund for their own billing bug is unconscionable and likely illegal — you cannot charge customers for services they didn't authorize
The initial support response refusing compensation appears to be AI-generated or macro-based, representing the dystopian future of customer service where chatbots are designed to deny legitimate claims
The existence of the HERMES.md detection code itself is deeply concerning — either someone deliberately coded string-matching billing logic or they are vibe-coding their billing system without proper review
Multiple users report independent billing issues (phantom invoices, random charges, triple billing) suggesting systemic problems beyond just this bug
The only reason Anthropic reversed course was public pressure on HN — without going viral, the original no-refund policy would have stood, which is a terrible model for customer support

Opposed

The support response may have been a poorly configured macro or decision tree rather than intentional malice — bad customer service routing is different from corporate fraud
Anthropic did ultimately fix the bug quickly, issue refunds, and provide bonus credits — the system worked once the right people were aware
Some commenters noted the support response may have been distinguishing between 'refund' (returning incorrect charges) and 'compensation' (additional payment for inconvenience), and was actually agreeing to a refund
Third-party harnesses like HERMES violate Anthropic's terms of service, so some detection mechanism is arguably justified even if the implementation was poor
People are overreacting — this is a growing startup that had a support escalation failure, not a deliberate scheme to defraud customers