top of page

OpenAI proposes perplexity limits on frontier AI development

55 minutes ago
8 min read
OpenAI proposes perplexity limits on frontier AI development, Data Studios original cover

OpenAI has published its internal pretraining perplexity scores for the first time and proposed that frontier AI developers agree on common limits, measured partly by perplexity, that they would commit to not crossing until the tools for understanding and monitoring AI behavior reach a concrete milestone. The proposal appears in a post titled "Perplexity and Intelligence", written by OpenAI chief scientist Jakub Pachocki and dated November 22, 2026, in the company's Research and Safety sections.


The post links specific loss values on OpenAI's main evaluation dataset to the capabilities that appeared at each level. GPT-3.5 sat at a log-perplexity of about 0.99, GPT-4 at about 0.91, and a model at about 0.76 began producing what OpenAI calls groundbreaking results on open mathematics problems the company had been posing to its models for more than a year. Pachocki writes that transformative capabilities in cybersecurity and mathematics emerged below 0.8, and that he expects further transformative capabilities, possibly ones nobody currently tests for, around 0.68 to 0.7.


The concrete request to the industry is twofold. Frontier developers should jointly set shared norms for evaluating models and red-teaming their safety cases before first use, including shared transparency standards, and they should accept common limits based on perplexity and other measurements. Pachocki gives one example of the milestone that would allow developers to move past a first limit: the development of a successor to chain-of-thought (CoT) monitors.


OpenAI presents the publication of its own numbers as the first step. The post acknowledges that the disclosure gives competitors a clearer picture of where OpenAI stands and says that concern is outweighed by the urgency of establishing shared safety measures across the industry. No publication schedule, no named partner and no date for a first limit appear in the post.


··········


OPENAI PERPLEXITY SCORES FROM GPT-3.5 TO THE 0.76 MODEL.


The figures below are the log-perplexity values OpenAI attributes to each capability step, all measured on the same held-out dataset drawn from OpenAI's internal codebase. The post names GPT-2 and GPT-3 as earlier milestones without giving their values on this dataset, and it does not name the model that reached about 0.76. The relative reduction column is Data Studios arithmetic on OpenAI's approximate values.


........


Model or threshold

Log-perplexity

Capability OpenAI links to it

Loss reduction vs previous step

GPT-2

Not given

First model to consistently form coherent paragraphs

n/a

GPT-3

Not given

Learning in context from a few examples

n/a

GPT-3.5

~0.99

Fluent conversation, basis of the original ChatGPT

n/a

GPT-4

~0.91

Productive chain-of-thought reasoning

8.1%

Below 0.8

<0.8

Transformative cybersecurity and math capabilities

n/a

Unnamed model

~0.76

Groundbreaking results on open math problems

16.5% vs GPT-4

Expected next level

~0.68–0.7

Further transformative capabilities expected

7.9%–10.5% vs 0.76


........


The distances look small on paper. The whole move from GPT-3.5 to the 0.76 model is a reduction of 0.23 in loss, or about 23% of GPT-3.5's value, and the step OpenAI now expects to bring another wave of capabilities is a further 0.06 to 0.08. Pachocki's argument depends on that compression of the scale: near the unknown floor of the loss, a small gain in prediction quality can correspond to a large jump in what a model can do once it is post-trained and prompted.


The GPT-4 entry carries a second detail. OpenAI writes that GPT-4's proficiency with chain-of-thought reasoning is what motivated the targeted elicitation of that capability and the creation of modern reasoning models, so OpenAI traces today's reasoning systems back to a capability first observed in a pretrained model. The math milestone connects to earlier claims the company has made about its models' research output, including the October announcement in which OpenAI said its models disproved the Exponential Time Hypothesis; the new post does not state which model or which results it has in mind.


··········


HOW OPENAI MEASURES PERPLEXITY ON ITS INTERNAL CODEBASE.


Perplexity measures how well a model predicts the next token of text it has never seen, and lower values mean better prediction. OpenAI says that in practice it tracks the log of perplexity, which is the same quantity as cross-entropy loss, and the values in the post are expressed that way. The primary held-out datasets OpenAI has used since GPT-3.5 come from its internal codebase, and the published numbers refer to a fixed snapshot of that code.


Pachocki frames pretraining as training a compressor, or equivalently a predictor, on a very large corpus, and stresses in a footnote that the word refers to a system able to compress by predicting unseen data. It does not mean a compressed archive of the training set, and the best compressor trained on a corpus does not necessarily remember its contents. Most of a modern model's capability, in this view, comes from that pretraining process, while prompting, fine-tuning and reinforcement learning elicit capabilities the pretrained model already holds in latent form.


The choice of a code dataset has a cost that OpenAI states openly. Given what GPT-2 and GPT-3 were trained on, their ability to predict code is a poor proxy for their general capabilities, so the early milestones are measured on a yardstick that favors newer models. Pachocki says he reported codebase perplexity because those are the numbers OpenAI relies on most today to gauge pretraining progress, and that joint evaluation of old and modern models on general text should still be informative.


··········


THE COMMON LIMITS PROPOSAL FOR FRONTIER AI DEVELOPERS.


The proposal has three layers. The first is a set of shared norms for how frontier models are evaluated and how their safety cases are red-teamed by the time the models are first used, together with shared transparency standards. The second is a commitment by developers to not cross common limits, defined by perplexity and other measurements, until the ability to understand and monitor AI behavior reaches a concrete milestone. The third is a framework for setting successive limits once the first one is cleared.


In Pachocki's words, the goal is an industry standard that specifies a first limit on development, measured by a key metric that can be tracked across labs, and states what developers would concretely need to achieve before going beyond it. The successor to CoT monitors is offered as an example of such a milestone. The post does not set the value of a first limit, does not say whether 0.68 to 0.7 on OpenAI's dataset would be a candidate, and does not describe who would verify compliance.


The case for acting now rests on the expected behavior of oversight tools. Pachocki writes that developers should expect further degradation in the currently dominant forms of alignment auditing and further increases in evaluation and training awareness, meaning models that behave differently when they detect they are being tested or trained. OpenAI has built much of its recent safety work on monitoring model reasoning, as described in our report on how OpenAI expanded misalignment monitoring to all frontier RL runs, and the new post treats a successor to those monitors as a precondition for pushing pretraining further.


The idea continues a line Pachocki took in September, when he called for an AI slowdown and mandatory safety bars. The November post adds a measurable variable to that position and OpenAI's own data points to anchor it.


··········


THE 2026 ACCELERATION IN OPENAI PRETRAINING LOSS.


A footnote states that the plot of losses of OpenAI models shows a steep acceleration in 2026. OpenAI attributes most of it to the outcomes of long-term research and compute investments and a smaller part to automation, adding that for the most recent models it has already seen meaningful design contributions coming from AI. The body of the post draws the forward-looking conclusion that as soon as 2027 the industry could conceivably see a similar avalanche of fundamental pretraining advances.


The post contains no compute figures, no training dates for individual models and no loss values beyond the milestones listed above, so the size of the 2026 acceleration can only be read from OpenAI's chart. AI-assisted design of AI systems is the same mechanism other developers have flagged in public documents, including the risk factor in which Anthropic's amended S-1 said a recursive self-improvement threshold may be imminent. OpenAI's wording is more limited, describing contributions to design and assigning the bulk of the acceleration to conventional research and compute.


··········


LIMITS OF PERPLEXITY AS A SHARED SAFETY METRIC.


Pachocki lists the weaknesses of his own proposal in some detail, and they shape how far the metric could be used across companies.


........


Challenge stated by OpenAI

Practical consequence

Labs use different datasets and training mixtures

Scores are not comparable across labs without a common test set

Perplexity does not map directly to tasks

Its value lies in predicting emergent capabilities

Chain-of-thought gains do not lower perplexity

Post-training gains escape a pretraining limit

The loss floor is unknown

Gains slow near the floor while capability jumps may grow

Code data is a poor proxy for older models

Dataset choice changes how milestones look


........


The first point is the hardest obstacle for an industry standard. A limit expressed as a log-perplexity value only binds competitors if they evaluate on the same held-out data, and OpenAI's figures come from a private codebase snapshot that other labs cannot use as published. The third point means a pretraining limit would leave reasoning, tool use and agent scaffolding outside its scope, which is why the post describes perplexity as one measurement among others.


··········


WHAT THE PERPLEXITY PROPOSAL CHANGES FOR CHATGPT USERS, API DEVELOPERS AND OTHER LABS.


The post announces no change to ChatGPT, Codex or the OpenAI API. There are no new models, prices, usage limits, plan changes or deprecations attached to it, and nothing in it alters current access for subscribers or developers. Its effect is on disclosure and on the terms of a possible industry agreement.


For developers and enterprise buyers who follow model roadmaps, the most concrete new information is OpenAI's own expectation that a loss level around 0.68 to 0.7 could bring further transformative capabilities, combined with the statement that a comparable wave of pretraining advances could arrive as soon as 2027. The post does not say when OpenAI expects to reach that level, and it gives no indication of whether OpenAI would pause before it absent an agreement with other labs.


For the other frontier developers, the proposal is an open invitation without a mechanism. The post records no commitment from any other company to publish perplexity figures or to accept common limits. It arrives while the largest labs face outside pressure on the same question, including the lawsuit in which OpenAI, Anthropic and Google were sued to freeze AI models at today's level, and it offers a voluntary, metric-based alternative to limits imposed from outside.


··········


OPENAI PERPLEXITY LIMITS: CONFIRMED FACTS AND OPEN QUESTIONS.


OpenAI has confirmed four things in its own words: the approximate log-perplexity values of GPT-3.5 (~0.99), GPT-4 (~0.91) and an unnamed model (~0.76) on its internal codebase dataset; the emergence of transformative cybersecurity and math capabilities below 0.8; the steep acceleration in its pretraining loss during 2026, mostly from research and compute; and the proposal of shared evaluation norms and common development limits tied to a monitoring milestone such as a successor to CoT monitors.


The open questions are the ones that would turn the idea into a standard. OpenAI has not named the model at 0.76, has not given the values of its newest models, has not proposed a numeric first limit, has not described a shared dataset that would make scores comparable across labs, and has not set a schedule for further disclosures. Until other developers respond with numbers of their own, the published scores describe OpenAI's trajectory alone, and the common limits remain a proposal from one company's chief scientist.


··········


FOLLOW US FOR MORE.


·····


DATA STUDIOS


·····


datastudios.org


bottom of page