Leaked 'Astra' Benchmark Scores Show Massive Leap in Math

teortaxesTex · x · 2026-08-30

Leaked benchmark scores for a model dubbed 'Astra' show impressive results on max settings: HLE (with tools) at 76%, CritPt at 45%, FrontierMath tier 4 at 97% (42/43), Riemann Bench at 100% (25/25), and OpenAI's internal FrontierScience Gold Set at 71%. A commentator notes that if real, this represents a substantial leap, given that the previous generation smaller pretrain 'Sol' was already 2/3 of the way there in raw scores.

Related event: Rumors mount around OpenAI's GPT-Astra with early-September launch tipped(18 posts)→

Original post →

More from Models

Models channel →