Crashing out at Anthropic and getting Pi pilled
TL;DR
Anthropic’s real problem isn’t just product mistakes — it’s a comms culture built for when everyone still loved them — Theo argues Anthropic still acts like criticism is propaganda, while OpenAI’s “how do we avoid disastrous comms?” mindset has helped it win developer sentiment.
The Claude Code chaos stacked up fast: rate-limit cuts, buggy-feeling usage collapses, a source-map leak, and a subscription policy reversal with ~18 hours’ notice — users paying $200/month suddenly found their workflows kneecapped, especially anyone using wrappers like OpenCode or OpenClaw.
The source-code leak was embarrassingly preventable — Ben and Theo say Claude Code releases were being published manually from team machines instead of CI, so stale source maps likely rode along in the NPM package and exposed the source.
Anthropic’s subsidy model may be the original sin behind all of this — Theo’s hottest take is that the $100/$200 Claude Code plans, which allegedly gave people 5x to 20x+ more inference than they paid for, were unsustainable marketing spend that warped the whole ecosystem.
Theo says Anthropic’s DMCA spree became its own fiasco — after the leak, GitHub repos were mass-flagged, including Theo’s fork of the official Claude Code repo with a one-line frontend prompt tweak, leading him to call Anthropic the most prolific GitHub DMCA sender and reverser.
The back half turns into a love letter to Pi, a tiny open-source coding agent that wins by doing less — unlike Claude Code’s giant prompts and tool bloat, Pi uses a minimal toolset, short system prompt, and self-customizing extension model that Ben says makes it feel cleaner and more powerful.
The Breakdown
A podcast debut built around crashing out
Theo and Ben open by making fun of podcast tropes, their folding-table setup, and the fact they’re hand-holding SM7Bs at 11 p.m. in Theo’s apartment. Theo frames the whole show as a place to talk about things that don’t fit into clickbait videos — basically, the offline conversations where he “breaks Ben’s brain,” now on camera.
Anthropic Week, and why Theo thinks the company would rather burn than admit he was right
They jump straight into “Anthropic Week,” listing the hits: Claude Code leak, subscription drama, awful comms, and policy confusion around using subs outside Anthropic’s own harnesses. Theo’s big read is that Anthropic built its public posture back when everyone loved Claude, and now that sentiment has flipped, the same strategy just looks arrogant and incompetent.
OpenAI wins points simply by not being miserable to deal with
One of the most revealing anecdotes is Theo texting an OpenAI friend a tweet saying OpenAI is the only lab that actually listens to feedback. Instead of celebrating, the friend asks, “Do you want me to clarify we’re not paying you?” and then worries about avoiding bad optics — which Theo uses to contrast OpenAI’s overcorrection with Anthropic’s refusal to communicate through official channels.
The rate-limit mess that made paying users feel gaslit
Ben walks through how users noticed Claude usage collapsing during peak hours, especially around 7 to 11 a.m. Pacific, with some people reportedly burning through usage in just a couple prompts. Theo’s real frustration is the timing: Anthropic announced the rate-limit change at 12:46 p.m., after the worst of it had already happened, and only through scattered employee posts rather than official Claude accounts.
Why caching matters, and why people thought Claude Code was secretly broken
Theo gives a pretty sharp explainer on inference caching: if anything in the prompt history changes, the cache can get blown and costs spike. Reddit users suspected a Claude Code bug was nuking cache hits, but the two land on a simpler explanation — Anthropic cut limits because it didn’t buy enough GPUs, and users felt the fallout while Anthropic charged unusually aggressive cache-write pricing like $6.25 per million tokens.
The source-map leak was less “elite hack” than “someone forgot to clean the folder”
This is the most technical and funniest part: Theo says Claude Code was being published to NPM directly from employee machines, not a CI pipeline. Their theory is that a local build with source maps left stale files in the dist folder, and because nobody wiped it before publishing, Anthropic accidentally shipped the map files that basically handed out the source.
Theo’s hottest take: the Claude Code subscription never should have existed like this
Theo argues the $100/$200 Claude Code subscriptions were massively subsidized — as much as 5x to 10x at first, now maybe 20x+ — and may have been introduced partly to lure Boris and Kat back after leaving for Cursor. In his view, that decision poisoned the ecosystem: once Anthropic offered “$5,000 for $200,” everybody else got dragged into trying to match an unsustainable deal.
The wrapper ban, Boris’s replies, and getting “Pi-pilled”
They roast Anthropic for emailing users with less than 24 hours’ notice that non-Claude-Code harnesses would lose subscription access, then getting tangled in contradictory replies about what wrappers are allowed. From there the mood flips: Ben gets animated about Pi, a tiny coding agent by Mario that uses just a few tools, a tiny system prompt, and a self-customizing extension system — the anti-Claude-Code philosophy, and the thing that clearly has him converted by the end.