Counterpoint from roon (OpenAI):

  last half of this article is poor analysis:
  - makes assertion that 5.2 token inefficiency ruins long horizon planning? and yet it tops the METR chart for long horizon planning? half baked
  - counts NPM downloads as authoritative when Claude code numbers hugely inflated bc GitHub actions does automatic CC download every time CI runs (vs Codex Compute cloud)
  - revenue numbers look like ragebait to me