From this screenshot of a chart given by OpenAI in their Navier-Stokes post, we can estimate that they have unreleased solutions to least 90 open problems. The pixel heights of the pass rates are separated by integer multiples of 1.698, and the 0.1 pass rate ticks are separated by 64.86 pixels, so N_attempts ≈ 648.6 pixels per unit pass rate/1.698 pixels ≈ 382 total attempts. If they used ±1 standard error for the bars, we can use the half width h of each bar (at its pass rate p) to estimate the number of problems N_problems=p(1-p)/h^2. The average for the 11 bars is 201. Consistent with both of our estimates, we assume that each problem was attempted twice, giving 382/2=191 problems. The height of the top dot corresponds to a 0.476 pass rate, so 191*0.476≈91 solved open problems.