Exponential View #600: GPT-6 Astra tops benchmarks but safety researchers doubt alignment claims

Exponential View (Azeem Azhar) · rss · 2026-09-06

Key points

Azeem Azhar's Exponential View hits its 600th weekly edition (11 years), with a limited-time 60% membership discount and an AMA.

Model watch

GPT-6 Astra beats Claude Fable 5.1 and other leading models on several benchmarks; the author's hands-on experience is highly positive and notes it's mostly cheaper than the Anthropic alternative. He's using it to organize 5,932 files on his desktop.
On hard math, Astra's time horizon is 30.9 minutes vs 3.6 for GPT 5.6 Sol; mathematician Bartosz Naskręcki says mathematicians can now focus purely on ideation and exploration.
ARC-AGI-3: Astra used fewer actions than the human median on 96% of completed levels, averaging 51.7% fewer actions per level — it invented a symbolic notation system to model entire environments, replacing costly trial-and-error with reasoning.
Safety controversy: researcher Ryan Greenblatt disputes OpenAI's "most-aligned model" claim, arguing that misaligned behaviors dropping from high rates in GPT 5.6 to zero in Astra looks like whack-a-mole patching rather than solving underlying misaligned drives.

Paid-member sections cover declining model interpretability, cancer vaccine manufacturing, and worker outcomes under advanced AI.

More from AGI Musings