GPT-5.4 sets new FrontierMath record

This is the kind of benchmark jump that keeps moving “frontier reasoning” from hype into measurable capability, but it also shows how fragile top-line scores can be when hard evals have limited sample sizes and possible shortcut paths.

• –FrontierMath is unusually high-signal because the problems are original, expert-written, and far harder than mainstream math leaderboards
• –GPT-5.4 Pro solving a never-before-solved Tier 4 problem is the standout detail, even more than the headline percentage
• –Epoch disclosed that OpenAI funded FrontierMath and has exclusive access to many problems and solutions, so the held-out analysis is important context rather than footnote material
• –The 38% Tier 4 pass@10 result suggests the model is getting meaningfully stronger with repeated attempts, which matters for agent-style workflows that can retry or branch
• –One newly solved problem appears to have been shortcut via a 2011 preprint, a reminder that benchmark wins still need careful interpretation before being treated as pure reasoning breakthroughs