All figures here are Theo's estimates or statements at recording time. They are not independently verified prices, allowances, or legal conclusions.

Reported capacity

  • $200 spent on either API buys $200 of metered usage; Theo argues personal subscriptions allow much more inference for internal coding work. 5:23
  • For a $200 Claude subscription, he estimates about $8,000 of monthly API-equivalent usage, with only $4,000 available to Fable because of its separate 50% allowance. 6:11
  • He rejects the $100 Claude tier for heavy use: it has half the total allowance but one quarter of the five-hour allowance, making the monthly capacity harder to consume. 6:31 6:51
  • For a $200 Codex subscription, he estimates about $12,000 of inference and says Astra and cheaper models such as Soul share the allowance without a separate model bucket. 6:51 7:07
  • He quotes Fable at $10 per million input tokens and $50 per million output tokens, compared with an earlier Mythos preview offer of $25 input and $125 output. 7:07 7:29 7:56
  • His explanation for low prices is competitive pressure and large margins. The roughly 95% margin figure is explicitly rumor and informal calculation, not disclosed company accounting. 8:12 8:31 8:53
  • He cites 14 Codex resets in 30 days and extrapolates potential usage to roughly $24,000. He also says new $200 subscriptions were temporarily paused. Both statements describe conditions at filming. 10:05
  • He favors direct Claude and Codex subscriptions over Cursor, OpenCode, or T3 Chat for this purpose, arguing third-party discounts are much smaller. His suggested 30–50% Cursor discount is speculation. 15:38 16:12 16:44

Privacy and permitted use as presented

  • Theo recommends disabling "Improve the model for everyone" under data controls and mentions a similar Claude setting. His assertion that this makes personal and team terms effectively identical is his claim, not a legal assessment supplied by these notes. 11:00 11:26
  • He draws a firm boundary around user-facing inference: use the API for public requests. He says subscriptions should support internal coding and experiments, not resale of inference or public traffic. 11:51 12:08
  • His later Opus 5.5 update reduces the need for this elaborate setup: he finds its behavior similar to Fable 5.1 and struggles to exhaust one subscription, so multiple accounts or a proxy may be unnecessary. 19:43 20:05