Astra Is Hard to Monitor
Zvi Mowshowitz @TheZvi · substack · 2026-09-08 · ★★★ · archived
The most detailed independent analysis of the Astra system card's CoT-monitorability findings; it was cross-posted to LessWrong and shared widely.
Summary
Zvi walks through OpenAI's own system-card evidence that Astra's chain of thought is much less monitorable than GPT-5.6 Sol's. He argues that capability gains explain only part of the drop and that architecture or training changes probably account for the rest. He highlights evidence that the model can shorten its reasoning when it knows it is being monitored, and quotes the card's warning that "we will lose confidence in our monitors" if the trend continues. He calls for industry coordination to avoid a race to the bottom on monitorability, and says a pause may be needed if monitoring cannot keep up. Also posted at thezvi.wordpress.com/2026/09/08/astra-is-hard-to-monitor/ and on LessWrong; shared on X (x.com/TheZvi/status/2097286192983073201). Date confirmed by fetching the Substack page.
Archived text
Page title: Astra Is Hard to Monitor
Page description: OpenAI’s central message on Astra is that it is three things:
Metadata archived 2026-09-29; see Summary for content.
Related events
All posts · id: 2026-09-08-zvi-astra-hard-to-monitor