Will OpenAI’s Astra model debut on the Arena Leaderboard at a score of at least 1490?

What this question means

Sixty-six percent, up from 56% a week ago, having ranged between 23% and 70% since tracking opened at 62%.

This asks whether a particular model debuts on the Arena Leaderboard at 1490 or above. Its companion asks the same at 1520 and sits at 8% — a fifty-eight point gap for thirty points of score.

That steepness is the interesting part. On most threshold pairs, a modest increase in the bar produces a modest fall in probability. Here a small step up the scale collapses the odds almost entirely, which says the crowd has a fairly precise expectation of where the score lands and considers 1520 well beyond it.

Sharp drops between adjacent thresholds are informative in a way gentle ones are not. They locate the edge of the crowd's expected range rather than merely sloping away from it.