GPT-5.5 Jumps to NYT Connections No. 2

This is a real benchmark win for GPT-5.5, but the more interesting signal is how crowded the frontier has become: one step down in reasoning effort can still move a model several points, and open weights are now close enough to matter operationally.

• –GPT-5.5 xhigh goes from 94.0 to 97.5, high from 93.6 to 96.9, medium from 92.0 to 95.0, and no reasoning from 32.8 to 37.5.
• –Gemini 3.1 Pro Preview remains #1 at 98.4, so GPT-5.5 is strong but not a new leader.
• –Kimi K2.6 at 91.4 is the standout open-weights result, ahead of Kimi K2.5 at 78.3 and well above DeepSeek V3.2 at 50.2.
• –Opus 4.7 looks weaker on this benchmark than Opus 4.6, especially with the high-reasoning refusal rate noted in the source thread.
• –This benchmark is a narrow reasoning test, so I would treat it as a model-selection signal, not a general verdict on coding or agent quality.