Theo Browne compares reported benchmark scores and token use, arguing that task-level cost depends on efficiency as well as the price per token. He explains how prompts, tool results and prior responses become input context, and how caching and compaction affect repeated work.
Theo Browne demonstrates reasoning token use on a small naming task and distinguishes visible answers from summaries of hidden reasoning. He describes examples of terse reasoning traces and speculates that compact internal reasoning may contribute to efficiency.
Theo Browne discusses possible tradeoffs involving long contexts, transparency and provider incentives. He presents these explanations as his own hypotheses rather than verified accounts of proprietary model internals.
Watch on YouTube
