Leaked scores put OpenAI internal models at 175-180 as safety takes back seat

scaling01 · x · 2026-11-18

scaling01 cites reports that OpenAI's internal models score between 175 and 180, which makes sense given Anthropic's last-gen model already scores 174. The more interesting launch will be the new Astra—the first generational upgrade of looping models—testing whether OpenAI cares about CoT monitorability or just frontier performance. Author's take: safety remains second priority at both Anthropic and OpenAI as capabilities dominate, especially with Chinese competitors closing in next year.

Related event: Rumors: Anthropic and OpenAI Both Have Stronger Models Ready(2 posts)→

Original post →

More from Models

Models channel →