Will OpenAI’s Astra model debut on the Arena Leaderboard at a score of at least 1520?

What this question means

Eight percent, drifting down from 13% at the start of tracking through a 2% to 13% band.

The pairing is what makes this readable. The same question at a 1490 threshold sits at 66%; at 1520, it is 8%. Thirty points of leaderboard score, and the probability falls by nearly sixty points.

A gap that steep means the crowd is not uncertain about roughly where the score lands — it is fairly confident, and 1520 sits outside the range it expects. Compare with a threshold series where each step down costs fifteen or twenty points; that pattern indicates a broad distribution. This one indicates a narrow one with a hard edge.

Benchmark questions have a useful property: they resolve against a published number on a public leaderboard, so there is no argument about the outcome. The only uncertainty is the score itself.