Opus 4.6 and Codex 5.3: The System Cards Matter More Than the Marketing

The article discusses the release of Opus 4.6 and Codex 5.3, highlighting the importance of the system cards over the marketing hype. It focuses on the capabilities and limitations of the models, particularly Codex 5.3's increased safety measures that can block CLI agents.

Details

The article discusses the simultaneous release of Opus 4.6 and Codex 5.3, and how the system cards are more interesting than the marketing. Opus 4.6 is positioned as the 'Architect', with improved capabilities in reading diffs, understanding git graphs, and multi-modal refinement. Codex 5.3, on the other hand, is the 'Builder' but has a higher confidence threshold for destructive commands, leading to increased safety refusals that can be a blocker for CLI agents running in trusted environments. The article suggests adjusting system prompts to provide the necessary authority context. The 'Atom everything' approach, where smaller, highly specialized sub-models are used instead of a monolith, validates the sub-agent architecture pattern. The article also highlights the importance of reading the system cards, as they reveal real limitations like 'over-refusal in shell environments' for Codex 5.3, which should be considered before integration.