Why Theo expects native combinations to improve
- He describes an earlier pattern where a model learned generic tool calling and harness developers adapted their execution systems around it. He says models are now increasingly trained on interactions from the harness itself.
43:1943:45 - In his account, Claude models generally behave better in Claude Code. He expects OpenAI models to benefit similarly from Codex, but criticizes Codex's current system prompt and ecosystem as an exception. These are his assessments at recording time.
43:45 - His prediction is that successful training and integration will make the native model-and-harness combination the natural choice, reducing the value of choosing the harness independently.
43:45
Numbers and evidence limits
- Theo estimates a roughly 3–5% difference, possibly negligible, between using Codex models through Claude Code and using them through Codex. He does not identify a metric, dataset, sample size, or controlled test in this passage.
44:40 - His reference to a competing harness's "3% win" is a rhetorical example involving older or open-weight models, not a benchmark report with enough detail to reproduce.
45:24 - He argues that small gains can be outweighed by subscription economics, token access, and total completed work. The video supplies no cost-per-task calculation for this comparison.
45:47
Decision rule
- Choose the capability needed at the price wanted. Theo's recommendation is to prioritize productive work over repeated debates about which harness is superior.
44:1145:47 - This is a claim of diminishing returns, not identical behavior: he acknowledges niche differences and criticizes particular prompts, while arguing that the remaining gaps deserve less attention.
39:0944:40