OpenAI just announced a batch of mathematical results produced by its internal frontier model: 722 manuscripts and 372 result families. @openai Is this using unsolved scientific problems as a new way to test models? Most problems were solved using the same process and a single agent, with an average computational cost roughly equivalent to ChatGPT Pro thinking for three hours. The significance of these weights in the scientific community isn’t as high as millennium problems like September’s Navier–Stokes. The heavier topics this time include Mahler, Kaplansky, free group factors, four-dimensional Kakeya, and advancements on the Riemann hypothesis. An average of three hours of reasoning time suggests it might be the same undisclosed frontier model, but using batch lightweight attempts—or possibly a distilled (flash) version of the model. Although OpenAI’s announcement was relatively low-key, each result family here might represent the equivalent of a top university math professor leading a team for several years, or even a decade, to achieve these outcomes. More importantly, similar to mathematics, many subfields with high formalization levels are likely already within the reach of OpenAI’s current models. Including but not limited to: - Statistical mechanics, quantum many-body systems, fluid dynamics, and dynamical systems in theoretical physics: existence, formulas, and phase transitions - Some theoretical chemistry and biophysics topics: spin glasses, mathematical models of protein folding, theorem-like propositions in molecular dynamics - Highly formalized engineering theories: bounds and convergence proofs in control, information theory, and optimization - Theoretical computer science problems that can be framed as conjectures or complexity propositions: algorithms, approximations, and reductions in cryptography And all of the above might just represent the level achievable by OpenAI’s weakest model in the next six months.