Put the note's own two numbers on one axis. Horizons have risen about 6x in the nine months since the paper [17], which works out to a doubling roughly every 3.5 months [18]. The published 95% interval on the current frontier measurement spans about 11x end to end [19], which is three and a half doublings, so at the trend's own rate the interval around a single model is as wide as roughly a year of progress [20]. A plan that says capability crosses some hour threshold in a named quarter is quoting a measurement that cannot resolve the quarter.
The domain spread is the part that should reach security planning. Translated onto the same doubling rate, the visual computer use gap is five to seven doublings, or roughly 18 to 23 months behind the software and research figure [21]. Attack chains that need a model to drive a screen, whether that is a console session or an operator UI, sit on the lagging branch. The physical case needs no benchmark: METR's author puts Claude 4.5 Sonnet's real-world coffee-making horizon at about two minutes [9].
The 50% is per task attempt, and when a model does succeed it is usually much faster than the human baseline [3]. The hours on the y-axis are human hours priced at a coin flip.
The soft spots in the level mostly point one way. Aggregating multiple successful baselines with the geometric mean rather than the arithmetic mean holds averages about 25% lower [16], and failed baselines were excluded rather than read as evidence that a task takes longer than the time the baseliner burned, which the author says would increase measured task lengths [15]. Baseline conventions alone could move the figure by more than 1.25x [13]. Pushing the other direction, the baseliner pool was skilled professionals in software engineering, machine learning and cybersecurity, and the author notes top engineers such as lab employees would be faster [14].
The rate survives all of this better than any single level does. Roughly 3.5 months per doubling is what the evidence carries; the position of any one model on that line is good to a factor of a few, and a factor of a few is the entire distance between one planning assumption and its alternative.
