Adjust for what the plan buys

  • Theo uses Adam Holter's analysis, which adjusts benchmark task prices for subscription discounts. The resulting gap between OpenAI and Anthropic is much smaller than their API task-price gap. 7:34 8:38
  • Soul max costs less than half as much per benchmark task as Opus 5.5 max after subscription adjustment. At high reasoning, Theo quotes roughly $0.04 for Soul versus $0.05 for Opus. 8:01
  • Theo generally avoids max, especially on Anthropic models. He often uses Soul x-high and Opus 5.5 high, which he describes as nearly identical in effective subscription pricing. 8:01 8:38

The cost-capability frontier

  • The frontier identifies the best value available at different capability levels. The transcript calls it the “PTO line.” 9:03
  • On API pricing, Theo says OpenAI occupies most of the frontier. Haiku briefly qualifies, despite costing roughly twice as much as Terra low; Opus qualifies at the highest capability level. 9:03 9:29
  • On one subscription-adjusted chart, Anthropic occupies most of the frontier, with narrow exceptions for Soul. Theo then corrects his initial chart interpretation: the more favorable OpenAI chart includes usage resets. A frontier comparison must therefore specify whether resets count. 9:52 10:21
  • Theo calls Opus 5.5 the best available model at recording time. That is his capability judgment, not a conclusion that every task or reasoning setting favors Opus. 9:29