GPT-6 Astra (high) hits 92.9% on WeirdML, matching Fable 5.1 (max)

zainhas · x · 2026-09-05

A WeirdML benchmark run shows GPT-6 Astra at the high setting scoring 92.9%, tying Fable 5.1 (max) for the top score and setting new individual highs on 6 of 17 tasks. Astra also writes notably less code than Sol and recent GPT models, though still more than Claude. Commenters suspect higher compute tiers (xhigh/max) won't move the needle much further, with fuller analysis to come.

Related event: OpenAI launches GPT-6 'Astra' amid AGI-era claims and wave of hands-on tests(19 posts)→

Original post →

More from Models

Models channel →