Playbook
Tasteful Skills
“Tasteful Skills” argues that the best agent skills are not documentation or best-practice lists.
Back to Podcast Digest
Seriously, Anthropic??
TL;DR
Anthropic allegedly billed users based on commit-message strings, not actual usage — Theo spotlights a case where a Claude Code Max 20x subscriber on the $200/month plan was charged another $200 because a git commit mentioned “Hermes MD,” even though they hadn’t exhausted their included usage.
The trigger appears to be Anthropic’s ‘third-party harness detection’ colliding with git history in the system prompt — Theo demos that an empty repo with only a commit message containing an OpenClaw-related string can flip Claude Code into paid overage mode when ‘extra usage’ is enabled.
Anthropic’s caching justification doesn’t hold up in Theo’s telling — he steelmans the argument that first-party tools might cache better, then counters with T3 Chat data showing Anthropic cache writes were about half of a roughly $40,000 monthly bill, and turning caching off barely changed costs.
Claude Code’s pricing and limits are generous on paper but confusing in practice — Theo says the $200 Max 20x plan can deliver about $500/week or roughly $2,000/month of API-equivalent inference on Opus (priced at $5 per million input tokens and $25 per million output), yet Anthropic still aggressively restricts third-party use.
The real issue, in Theo’s view, is product strategy turning into user-hostile engineering — instead of warning users that tools like OpenClaw may burn usage faster, he argues Anthropic built brittle detection logic that punishes strings in prompts and commits, creating a bug class that suggests the whole approach is broken.
Theo frames this as a culture problem, not a one-off bug — he contrasts Dario Amodei’s near-silence on Claude Code with Sam Altman’s constant Codex promotion, and argues Anthropic’s engineering culture treats developers and users with open disdain.
The Breakdown
Theo opens in full crash-out mode
Theo says he’s tired of making videos about Anthropic, but they “keep finding new ways to embarrass themselves.” He starts from prior grievances — charging differently based on system prompts or unsupported tools like OpenClaw — then says the new low is charging based on filenames and even git commit text.
Why the $200 Claude Code plan matters
He explains the Max 20x Claude Code plan: $200/month, with what he describes as extremely generous included usage, roughly equivalent to about $500 of inference per week or $2,000 per month if you really push it. He notes Opus API pricing at $5 per million input tokens and $25 per million output, and says Anthropic profits from subscribers like him who barely use the allowance — which is exactly why they don’t want people routing that usage through outside tools.
Theo steelmans the caching argument, then tears it apart
To be fair, he walks through why Anthropic might care about third-party harnesses: poor caching can make the same workload much more expensive. He uses a “hibernation on a computer” analogy for prompt caching, saying first-party Claude Code should, in theory, cache better than generic tools like OpenClaw or OpenCode. But then he says Anthropic undercuts its own case by shrinking cache duration from an hour to 5 minutes and repeatedly breaking caching anyway.
T3 Chat’s $40,000 Anthropic bill is the receipt
Theo pulls in data from T3 Chat, where Anthropic models processed about 11 billion input tokens and 650 million output tokens in April, for a monthly bill around $40,000. The wild part: on some days, prompt cache writes alone cost $970 versus $300 for normal input and $500 for output, and after his team disabled caching on April 20, costs barely changed. His conclusion is blunt: Anthropic can’t credibly blame third-party tools for bad caching when their own caching “doesn’t work.”
The -p flag, system prompts, and the trapdoor to overage billing
He shows how Claude Code’s built-in -p mode lets you pass prompts programmatically — exactly the kind of CLI behavior wrappers and agents need. Add an appended system prompt mentioning OpenClaw, and Claude starts throwing an “out of extra usage” API error; with the dashboard’s ‘extra usage’ toggle on, those requests quietly become billable overages even before normal subscription limits are hit.
The absurd demo: a commit message alone can trigger billing
Theo then does the thing he clearly expects not to work: he creates an empty repo, makes a commit whose message contains the OpenClaw detection string, changes nothing else, and runs Claude Code. It still trips the same behavior. That’s the sticky, ridiculous centerpiece of the video — not a malicious exploit, just a string in git history being enough to push a user into paid usage.
Hermes MD and Anthropic’s public admission
He ties the demo to the real user report: a commit mentioning Hermes.mmd — a spec file associated with Hermes Agent — got pulled into Claude Code’s system prompt via recent git history and flagged by Anthropic’s third-party harness detection. Thor from Anthropic replied publicly that it was “a bug” caused by harness detection combining poorly with git status, and said affected users would be refunded and given another month of credits. Theo’s response is that this is the kind of bug that means the whole design is rotten.
From bug report to culture indictment
The final stretch stops being about one bug and becomes an all-out attack on Anthropic’s engineering culture and leadership. Theo argues Dario Amodei barely acknowledges Claude Code at all, contrasts that with Sam Altman talking constantly about Codex, and claims the company’s product and infra failures — including T3 Chat being unable to use Anthropic models because Anthropic’s own rebilling failed — reflect deep contempt for engineers and users alike. He ends exactly where the title promises: not with a fix, but with a furious “you’ve made me an enemy.”
Share
Keep Reading
Two Ways to Follow
The Weekly Echo. The inbox-shaped summary of what mattered.
New editorials announced here.
