Quit Yapping
Be careful what you name your markdown files...
34:40
Watch on YouTube ↗
T
Theo - t3.gg·Tech

Be careful what you name your markdown files...

TL;DR

Anthropic accidentally billed a user $200 because their third-party harness detection scanned git commit messages for banned tool names like 'Hermes.md'.

Key Points

  • 1.A user was billed $200 erroneously due to a commit message containing 'Hermes.md'. The user had not exceeded any usage limits — simply mentioning the filename in a git commit triggered Anthropic's third-party harness detection, which routed them to overage billing.
  • 2.Anthropic's Claude Code scans system prompts and git history to detect banned third-party tools like OpenClaw and Hermes Agent. This 'third-party harness detection' routes users to extra billing if disliked strings are found, even when usage limits haven't been reached.
  • 3.The host live-demonstrated the bug by creating an empty repo with a commit message matching the OpenClaw schema string. Without changing any system prompt, calling claude -p in that repo triggered the billing behavior, confirming the exploit required no malicious action.
  • 4.Anthropic's caching is demonstrably broken and more expensive than not caching at all. T3 Chat's Anthropic bill showed prompt cache writes consuming ~$970 on a day where input tokens cost $300 and output cost $500; stopping caching entirely did not increase costs.
  • 5.The $200/month Claude Code Max plan offers up to $2,000/month of inference at API rates, creating a 10x subsidy Anthropic protects by locking out efficient third-party tools. Anthropic profits from subscribers who underuse their plans, like the host who estimates he uses only $50/month worth.
  • 6.Hermes Agent is an open-source, minimal alternative to OpenClaw that also used the claude -p flag to wrap Claude Code programmatically. Anthropic attempted to ban it via the same system-prompt string-matching mechanism used against OpenClaw, leading to the erroneous billing bug.
  • 7.Anthropic's own API reliability failures locked T3 Chat users out of Anthropic models after an automatic rebill API call failed. Anthropic then sent emails claiming T3 Chat hadn't paid, despite the failure being on Anthropic's billing infrastructure side.
  • 8.The host attributes the root cause to CEO Dario Amodei's disdain for software engineers, which he claims originated from Dario's discomfort working under engineer-led leadership at OpenAI. He argues this cultural contempt cascades into Claude Code's team mistreating end users who rely on the product.

Life's too short for long videos.

Summarize any YouTube video in seconds.

Quit Yapping — Try it Free →