OpenAI's GPT-5 Pro set a new record on the FrontierMath Tier 4 benchmark by solving one additional research-level math problem compared to Google's Gemini 2.5 Deep Think. The evaluation involved manually grading three compute-intensive model settings on 50 extremely hard problems developed by professional mathematicians.
- GPT-5 Pro solved six problems (13%) via its API and the same number via its web app, achieving a combined pass@2 of eight problems (17%).
- Gemini 2.5 Deep Think was edged out by a single problem, a difference noted as not statistically significant.
- Grok 4 Heavy lagged behind the other two models in performance.
- The total number of Tier 4 problems ever solved increased to nine (19% of the benchmark), with five of GPT-5 Pro's eight solved problems coming from the held-out set.
This result highlights the increasing capability of frontier models to tackle research-level mathematics, with GPT-5 Pro demonstrating a slight edge over its closest competitor.