OpenAI internal model cracks Millennium math problem in 88 hours

This is a signal that frontier AI capability growth is no longer linear with public model generations, as the internal model's leap suggests deliberate withholding of top-tier performance. For the research and quantitative sectors, it implies that the hardest open problems may become testbeds for evaluating hidden model strength, not just benchmarks. The near-48% pass rate on selected problems, versus Astra's best, highlights how far ahead internal systems may be, which could shift expectations around what public releases represent. Watch whether OpenAI formalizes this as a new evaluation tier or if other labs respond with comparable internal disclosures, as that would reset the competitive baseline.