A benchmark number means nothing without the conditions it was taken under
The same model on the same benchmark turns up with different numbers depending on where you read them. That is not always an error. Usually it means the runs were not the same run.
Three things move the figure most. How many trials were taken and which is reported, whether a single pass, an average over several, or the best of several. How much the model was allowed to think. And how the model was wired into the task: which tools it could reach, what time limit it ran under, what it was permitted to do.
The SWE-bench Verified leaderboard states this for its own entries[1]. Claude Mythos 5 at 95.5 percent is an average over five trials at maximum thinking effort, and the page adds that environment, tool permissions, time limits and the surrounding machinery all matter[1].
The consequence is that a table of scores gathered from different announcements is not a comparison. It is a list. On this site a score is published only with the conditions it was taken under, and where two sources disagree both are shown rather than averaged into one.