Be careful what you name your markdown files...
TL;DR
Anthropic accidentally billed users based on Git commit text, not actual usage — Theo shows a real case where a Claude Code Max 20x subscriber on the $200/month plan got charged another $200 because a commit message contained the string “Hermes.md,” which Anthropic later called a bug in its “third-party harness detection.”
The trigger appears absurdly broad: even a commit message can flip Claude Code into paid overage mode — In Theo’s demo, an empty repo with no special system prompt still tripped billing behavior after he made a commit containing an OpenClaw-related string, suggesting Anthropic scans prompt context assembled from Git history and routes requests differently if it sees banned tooling markers.
Anthropic’s caching justification doesn’t hold up against Theo’s own bill — He tries to steelman the company’s argument that third-party tools like OpenClaw cost more because they cache poorly, then points to T3 Chat’s April Anthropic bill: about $40,000 total, with prompt cache writes consuming roughly half on some days, including one day with $970 in cache-write charges versus $300 input and $500 output.
Claude Code’s subscription economics explain the lock-in incentives — Theo says the $200 Max 20x plan can effectively deliver around $500/week or roughly $2,000/month of Opus inference at API-equivalent pricing, so Anthropic benefits when subscribers underuse it and has an incentive to block programmatic or external harnesses that would push customers closer to that ceiling.
The core issue isn’t just a bug, it’s the product strategy behind it — Theo argues that if “third-party harness detection” can collide with Git history in the system prompt and bill users for innocent strings, that’s evidence the whole anti-harness enforcement path is rotten, not merely a one-off implementation mistake.
Theo contrasts Anthropic with OpenAI using his own spending, not hypotheticals — He says T3 Chat uses OpenAI models two to four times more than Anthropic models while the OpenAI bill is only a bit over half as much, and he frames his criticism as coming from a customer spending about $40,000/month on Anthropic rather than a rival taking cheap shots.
The Breakdown
Theo opens with a fresh Anthropic horror story
Theo starts already exasperated: Anthropic, he says, has moved from billing weirdly based on system prompts and unsupported tools to effectively billing based on filenames and even Git commit messages. The headline example is a Claude Code user on the $200/month Max 20x plan getting hit for another $200 despite supposedly having unused subscription capacity, just because a commit included “Hermes MD.”
Why the $200 Claude Code plan creates weird incentives
He explains the economics of Claude Code’s top-tier subscription: at API list prices, a heavy user can squeeze roughly $500 of inference per week, or around $2,000 a month, out of the $200 plan. Theo’s point is that Anthropic wins when customers like him underuse the plan, and loses when people route that same subscription through agents or automations like OpenClaw that drain it efficiently.
Theo steelmans the caching argument — then tears it apart
To be fair, he lays out Anthropic’s likely argument in detail: first-party tooling should cache better than generic third-party harnesses, and better caching means lower compute costs and faster responses. He uses a hibernation analogy for model state, then says that if Anthropic really cared about helping users preserve cache, they wouldn’t quietly shrink cache duration from one hour to five minutes and constantly break their own system.
The T3 Chat receipts: caching costs are out of control
Then he brings numbers. T3 Chat, his product, processed about 11 billion Anthropic input tokens and 650 million output tokens in April so far, for a bill around $40,000; meanwhile OpenAI is used two to four times more and still costs only a bit over half as much. The killer detail is cache writes: on one day, Theo says, they cost $970 versus $300 for inputs and $500 for outputs, and when his team turned caching off on April 20, the bill barely changed.
The claude -p loophole and Anthropic’s anti-harness enforcement
Theo then walks through the key product mechanic: Claude Code supports claude -p, which lets you pass prompts programmatically from the CLI, plus appended system prompts. That makes it straightforward for tools like OpenClaw or Hermes Agent to wrap Claude Code — and, according to Theo, Anthropic began detecting banned strings inside those prompts and forcing requests into “extra usage” overage billing if the user had that switch enabled.
The demo gets worse: a Git commit alone can trigger it
The video’s wildest moment is Theo reproducing the issue in what he clearly expected to be a failed demo. In a brand-new empty repo, he creates a trivial Markdown file and makes a commit containing an OpenClaw-related schema string; with no special system prompt at all, just running claude -p is enough to trigger the bad behavior. His reaction is basically disbelief: the only explanation, he says, is that Anthropic is scanning Git-derived context inside the system prompt and matching on raw strings far too naively.
Anthropic confirms the bug — but also confirms the bigger accusation
Theo cites Anthropic engineer Thor’s response: “Sorry, this was a bug with the third party harness detection,” caused by an interaction with how Git status/history gets pulled into the system prompt. Refunds plus an extra month of credits were offered, but Theo says that only proves his larger point — if unrelated systems like harness detection and Git history can collide this badly, the entire feature path should be reconsidered, not patched.
The rant expands from bug report to company indictment
In the final stretch, Theo turns from product criticism to culture and leadership, arguing Dario Amodei doesn’t value developer-facing product work the way Sam Altman publicly celebrates Codex. He closes by contrasting OpenAI’s responsiveness with Anthropic’s silence, mentioning even a billing outage that locked T3 Chat users out of Anthropic models, and ends in full burn-it-down mode: this isn’t just bad code, in his view, it’s a company treating users and engineers with contempt.