GPT-5.4 tops ZeroBench leaderboard

This is a useful benchmark win because ZeroBench is still brutally hard, so even small gains usually reflect real progress in multimodal reasoning rather than leaderboard noise.

• –ZeroBench was introduced as an “impossible” visual benchmark, and frontier models are only now starting to post non-trivial scores
• –GPT-5.4 taking the top spot over Gemini 3.1 Pro suggests OpenAI is still highly competitive on image-heavy reasoning, not just text benchmarks
• –The absolute scores remain low, which is the bigger story for developers: multimodal reasoning is improving, but it is nowhere near solved
• –ZeroBench’s latest site update says its recent v3 wording tweaks did not affect scores, so this looks like a genuine model improvement rather than a benchmark reset