AI coding agents: Codex vs Claude, benchmarks and real use

A developer posted a week-long comparison of using OpenAI Codex more than Claude for real coding work, and the thread filled with nuanced takes. The positive: Codex produced Ruby/Rails code with fewer unnecessary comments, which the author preferred. The negative: multiple commenters flagged that Codex and similar 'sol' models tend to overengineer and be overly cautious, adding abstractions and error handling layers that weren't asked for. One commenter tried using it for scraping and data processing and found it kept building in hedges.

Separately, the NanoGPT Speedrun Frontier post measured 18 frontier models running 153 autonomous research tasks, specifically optimizing the nanoGPT training loop. This is a rare benchmark that tests agentic research capability, not just code completion. The results matter because the models are being asked to do genuine optimization work, not just transcribe instructions.

The pattern: the 'which AI coding agent' conversation is moving from vibes to structured comparison, but the comparisons are still noisy because the harness setup matters as much as the model. Commenters on the Codex thread noted the comparison was not useful without specifying exact model versions.