HorizonMath finds GPT-5.4 Pro gains on unsolved math

This is the kind of benchmark result that matters more than leaderboard gaming, but it should be treated as a strong signal, not a solved milestone, until peer mathematicians fully verify the proofs and constants.

• –HorizonMath targets open problems where verification is tractable, which makes it harder to fake progress with pattern matching.
• –Two concrete improvements on unsolved classes is notable because most frontier models reportedly score near zero on this benchmark.
• –The one-hour reasoning runtime hints that breakthrough-style outputs may be possible with longer test-time compute, not just bigger pretraining.
• –If expert review confirms both results, AI evals may shift toward “novel contribution rate” instead of only accuracy on known-answer sets.