GPT-5.5 benchmarks edge GPT-5.4

GPT-5.5 looks like a serious upgrade for real-world agent workflows, but the charts read more like a disciplined frontier optimization pass than a shock-and-awe leap. On OpenAI’s published evals, GPT-5.5 posts 82.7% on Terminal-Bench 2.0, 84.9% on GDPval, 78.7% on OSWorld-Verified, and 81.8% on CyberGym, beating GPT-5.4 on each. The coding story is strongest in tool-heavy, long-horizon tasks: OpenAI says GPT-5.5 also reaches 58.6% on SWE-Bench Pro and performs better than GPT-5.4 while using fewer tokens. It is not a clean sweep over every rival on every benchmark; Claude Opus 4.7 still tops GPT-5.5 on BrowseComp in OpenAI’s own chart, which keeps the frontier race competitive. The real developer takeaway is economics plus reliability: matching GPT-5.4 latency while needing fewer tokens is exactly the kind of improvement that makes agents more usable in production. Reddit reaction is already split between “solid upgrade” and “underwhelming relative to hype,” which usually means the benchmarks are good enough to matter but not dramatic enough to settle the model wars.