On November 22, 2026, OpenAI chief scientist Jakub Pachocki published an essay titled "Perplexity and Intelligence" on the company's website, disclosing for the first time the internal pretraining perplexity curve OpenAI has tracked for years. He proposed perplexity as a shared yardstick for pretraining across the industry, and called for common limits on frontier model development until the ability to monitor these models catches up.
Perplexity measures how accurately a model predicts the next token; the lower the number, the better the prediction. Pachocki treats it as equivalent to cross-entropy loss. The published curve uses a held-out dataset drawn from OpenAI's internal codebase: GPT-3.5 sits at a log-perplexity of about 0.99, GPT-4 at about 0.91, and the latest models have fallen below 0.8. The curve drops noticeably faster through 2026, and the chart also plots an unreleased internal model lower still.
What the number 0.76 means
Behind the numbers is a jump in capability. Pachocki writes that GPT-3.5, at about 0.99, could hold a fluent conversation, and GPT-4, at about 0.91, became adept at using chain-of-thought reasoning. Once perplexity fell below 0.8, transformative capabilities emerged in cybersecurity and mathematics. After moving to a model with a log-perplexity of about 0.76, OpenAI started seeing long-standing open math problems solved one after another. In the same week, the other big math story was the Lean community formally confirming the Riemann hypothesis proof produced with Claude's help. Taken together, the two events suggest mathematics is where frontier models are breaking through first.
Can one line hold back frontier models?
Pachocki is open about perplexity's weaknesses: labs use different datasets, so the numbers are hard to compare directly; the metric does not map onto any practical task, and it says nothing about the capability gains that chain-of-thought reasoning brings; and nobody knows where its theoretical floor lies. Even so, he argues it remains the single metric that best tracks progress across model generations. His concrete proposal is that labs agree on an industry standard setting a first shared limit, and only cross it once the ability to understand and monitor models reaches a specific milestone, such as a successor to chain-of-thought monitoring.
He also offers a timeline: a wave of fundamental pretraining breakthroughs could arrive as soon as 2027, and as models enter the 0.68 to 0.7 log-perplexity range, further capabilities that cannot be foreseen today should be expected, while alignment auditing weakens further and models grow more aware of being evaluated. The essay names no specific cap and says nothing about who would enforce one. For now it is one lab chief scientist's appeal; whether other labs take it up will become clear in the coming weeks.