Claude review boosts Codex accuracy by 18% in cross-model study

A new paper accepted at KDD'26's Agentic SE workshop provides concrete evidence on how to best pair coding agents. Researchers tested Anthropic's Claude and OpenAI's Codex on 116 recent hard and medium LiveCodeBench tasks under six conditions: solo baselines for each model, same-model self-review, and cross-model code review (Claude reviewing Codex, and Codex reviewing Claude). The reviewer saw the problem statement and the draft code but could not execute tests—mimicking a real peer code review step.

The results are strikingly asymmetric. When Claude reviewed Codex drafts, the pass rate jumped from 71.6% to 89.7% (p=0.001). Codex self-review also helped, but only to 84.5% (p=0.022). The reverse pairing backfired: Codex reviewing Claude drafts dropped the baseline from 91.4% down to 82.8% (p=0.046), mostly due to false positives. Claude self-review left the 91.4% baseline unchanged. The practical takeaway is clear: if you're using two LLMs for code generation, let Claude do the reviewing and Codex do the writing—not the other way around.