GPT-5.4 tops Extended Connections benchmark

GPT-5.4 looks meaningfully better on a puzzle-heavy reasoning benchmark, but the split between reasoning and no-reasoning modes is the real story for developers evaluating cost, latency, and capability tradeoffs.

• –The medium-mode jump from 71.4 to 92.0 is huge and suggests OpenAI improved practical reasoning efficiency, not just max-effort performance.
• –The benchmark uses 759 NYT Connections puzzles with extra trick words, so it is testing categorization and distractor resistance rather than straight factual recall.
• –The weak no-reasoning score relative to reasoning modes reinforces how much structured inference still matters on deceptively simple language tasks.
• –This is a strong directional signal, but it is still a niche third-party benchmark rather than a full proxy for coding, agent reliability, or production workloads.