We can expect a reliable PhD-level mathematician AI by early 2027, or by 2029 if the graph is logistic. Aggregating the three most important and diverse math benchmarks, from arXiv proofs to verification to long-horizon open problems, into one Math Perf value, 2027 looks to be the year when AI should theoretically be competent enough to tackle any math research problem. At that point, success may simply become a function of running your agent for a longer amount of time.
This doesn’t account for creativity or the ability to create totally new concepts, but I’d argue the vast majority of mathematicians are not creating “totally new concepts” either.