GPT-6 Astra reportedly solves competition math without verbalized reasoning, with sharply worse monitorability

MattGarciaEth · x · 2026-09-04

Ryan Greenblatt highlights benchmark results suggesting GPT-6 Astra is a massive jump in opaque reasoning: it appears to solve hard competition math problems entirely "in its head," without verbalized reasoning, while prior models handled only basic word problems. He flags benchmark caveats — especially contamination — as a key concern.

UK AISI reportedly found Astra has much worse monitorability, making internal reasoning harder to audit. Greenblatt guesses the jump comes from architectural changes with increased serial depth, though a normal large-scale pretrain scale-up is plausible; similar jumps in future generations would be deeply concerning for oversight.

Related event: GPT-6 Astra deemed more aligned but far less monitorable, alarming safety researchers(7 posts)→

Original post →

More from Models

Models channel →