GPT-5.4 writes the cleanest code of any GPT model we've tested — and fails more tests than 5.3. Erosion drops, duplication drops, but pass rate and core both regress. We dig into why.

Anthropic's Opus 4.6 copy-pastes. OpenAI's GPT-5.3 Codex over-abstracts. Both miss edge cases at the same rate. New SCBench results with a guide for when to trust each model.

AI coding agents become lazy patchers under iterative changes, copying code instead of refactoring. This creates massive god functions that are unmaintainable—explaining the gap between benchmark scores and real-world experience.

SlopCodeBench evaluates AI coding agents under iterative specification updates. Unlike single-shot benchmarks, SCBench reveals verbosity and structural erosion that make agent-written code unmaintainable over time.

Outcome Reward Models for code verification allow one to trade accuracy for speed in the generate-then-rank paradigm. This can be further improved through a generate-prune-then-rank approach where a weaker verifier prunes solutions prior to ranking, thus saving work on incorrect tokens. We show that this hybrid approach can be 11.65 times faster than running the whole test suite while only being 8.33% less accurate.