Crashing out at Anthropic and getting Pi pilled

TL;DR

Theo and Ben break down Anthropic's chaotic week of rate limit changes, source code leaks, and subscription bans that eroded developer trust.

Key Points

• 1.Anthropic's comm strategy assumed universal goodwill that no longer exists. Employees still believe anti-Anthropic sentiment is paid propaganda or temporary, a delusion that is accelerating their reputational collapse.
• 2.Anthropic intentionally reduced rate limits during peak hours (7–11 a.m. Pacific) without notifying users until 12:46 p.m. — two hours after the change ended. Users had no official way to know what was happening unless they followed third-party accounts like Thoric.
• 3.A suspected caching bug compounded the rate limit crisis. Anthropic uniquely charges $6.25 per million tokens for cache writes, while OpenAI charges nothing for cache writes and does automatic caching.
• 4.The Claude Code source code leaked because Anthropic never set up a CI/CD pipeline for publishing — every npm release was run manually from a team member's local machine. A dev build with source maps was left in the dist folder and accidentally shipped.
• 5.The leak revealed that Anthropic's 'secret sauce' code contained copy-pasted snippets from open-source projects like Open Code, not just inspiration. Boris originally called Claude Code their competitive secret that they feared giving competitors.
• 6.Theo argues the subsidized subscription model ($200/month for up to $5,000 in inference) was a mistake made solely to lure Boris and Cat back after they briefly left for Cursor. Every subsequent PR disaster traces back to that original economic decision.
• 7.Anthropic gave less than 24 hours' notice before cutting off all non-Claude-Code subscription usage, announcing the change on a Friday night after business hours with only $20–$200 in API credits as compensation.
• 8.Claude Code ranks 12th out of available harnesses for Opus model performance — meaning users who use any other tool get better code output, yet Anthropic only subsidizes inference through their own harness.
• 9.Boris confirmed via Twitter reply that custom wrappers for Claude Code (like T3 Chat/T3 Code) are allowed for local personal use, but Anthropic's official account walked it back, saying it wasn't guidance, leaving tool builders in legal and operational limbo.
• 10.OpenAI's positive sentiment shift is attributed not to payments but to a deliberate strategy of owning mistakes and responding to feedback, contrasted with Anthropic's pattern of never posting negative news on official channels.
• 11.The 'Pi pilled' element refers to Theo becoming increasingly convinced by PI/alternative AI tooling, with the duo teasing future nerd-snipe segments and expressing frustration that frontier Chinese labs (Moonshot, Zhipu) are pleasant to work with but not yet competitive enough to matter.