Sam Altman calls the post the most important thing OpenAI has published in some time and endorses collective industry adoption of common limits.
OpenAI Proposes Perplexity as Shared Industry Safety Threshold
The proposal asks rival AI labs to coordinate on a measurable development ceiling, with progress beyond it gated on demonstrable advances in AI interpretability. OpenAI is disclosing its own pretraining position to push for that coordination, accepting competitive exposure as the cost.
When stories like this move, I write about them in an email each morning.
One free email each morning, around 8 AM US Eastern.
Check your email.
I have sent you a confirmation link. One click and you are in.
Not in your inbox? Look in spam or promotions. If it's there, mark it "Not spam" so future editions land in your inbox.
The full picture
OpenAI chief scientist Jakub Pachocki published a blog post and perplexity chart on November 22, 2026, disclosing log-perplexity values for OpenAI's models from GPT-3.5 through GPT-6.2 Astra, including an unnamed internal model at approximately 0.73 and a shaded uncertainty band below roughly 0.71. The post argues that pretraining perplexity has reliably tracked major capability shifts: GPT-3.5 at approximately 0.99 enabled fluent conversation, GPT-4 at approximately 0.91 enabled productive chain-of-thought reasoning, and models below 0.8 showed what the post describes as transformative capabilities in cybersecurity and mathematics. Pachocki proposes that frontier labs agree on a first hard perplexity-based limit beyond which no lab would advance until concrete alignment milestones, such as a successor to chain-of-thought monitors, are met. OpenAI released its own internal measurements alongside the proposal, acknowledging competitive risk. The post also flags challenges with perplexity as a shared metric: cross-lab comparability requires a shared test set, the metric can detach from task-specific performance, and an unknown entropy floor means smaller perplexity gains may correspond to larger intelligence leaps. Sam Altman called this the most important thing OpenAI has published in some time and said the industry would be wise to adopt common limits collectively.
How it developed
Sources
- OpenAI Astra's chain-of-thought opacity
- Pachocki, UN Official, Journalist Flag AI Alignment Gaps Publicly
- OpenAI's AI research agents and self-improvement reporting
- OpenAI's proposal for binding global AI standards
BackgroundWhen AI doesn't do what we meant
Get the daily email
When stories like this move, I write about them in an email each morning.
One free email each morning, around 8 AM US Eastern.
Check your email.
I have sent you a confirmation link. One click and you are in.
Not in your inbox? Look in spam or promotions. If it's there, mark it "Not spam" so future editions land in your inbox.