GPT-5.5-Codex vs 5.3: A 200-Task Bench Result

On a 200-task bench split across a TypeScript SaaS and a Python ML pipeline, GPT-5.5-Codex closed 81% of tasks unattended versus 67% for GPT-5.3-Codex, and burned 38% fewer reasoning tokens on the multi-step ones. But on trivial single-file edits it ...