OpenAI Chief Scientist Releases Internal Perplexity Data, Proposes a Single Metric to Set a Global AI Safety Speed Limit

OpenAI Chief Scientist Jakub Pachocki has disclosed internal log-perplexity figures for successive models from GPT-3.5 onward and proposed that frontier la

On November 22, 2026, OpenAI Chief Scientist Jakub Pachocki published an article titled “Perplexity and Intelligence” on the company’s official blog, disclosing for the first time internal log-perplexity data for its successive models from GPT-3.5 to the present and proposing that frontier AI labs use this metric as the basis for establishing common industry safety standards. OpenAI CEO Sam Altman called it “one of the most important things OpenAI has ever published.”

The figures Pachocki released provide a complete capability-evolution curve: GPT-3.5 had a log perplexity of about 0.99, forming the basis of the original ChatGPT; GPT-4 was about 0.91, a level sufficient for models to put chain-of-thought reasoning to practical use, which in turn gave rise to OpenAI’s reasoning-model product line; the latest models have fallen below 0.8, and according to OpenAI, “transformative capabilities” emerged in cybersecurity and mathematics within this range, and no one had anticipated that these capabilities would appear at this level. When log perplexity reaches about 0.76, models begin to produce “breakthrough results” on open problems in mathematics.

Compression Capacity Is the Foundation of Intelligence

To understand the weight of this proposal, one must first clarify what perplexity measures. At the technical level, perplexity is equivalent to cross-entropy loss and measures the average accuracy with which a model predicts the next token—the lower the value, the more precise the prediction. Pachocki describes the core training process of modern AI as training a “compressor”: the ability to compress information across massive amounts of unstructured data constitutes the source of underlying intelligence.

This perspective breaks with the usual evaluation framework. Common AI benchmarks measure the “surface” of capability after fine-tuning and reinforcement learning. Perplexity, by contrast, measures a deeper “intelligence substrate” from the pretraining stage: how strong the model’s ability is to compress and predict linguistic information itself. Pachocki draws a clear distinction between two things: first build a “brain” with latent capabilities through pretraining, then elicit its potential through prompt engineering, fine-tuning, or reinforcement learning. Pretraining perplexity is the core measure of the former.

Judging from the historical data OpenAI disclosed, the correlation between this metric and capability emergence is quite stable. GPT-2 reached a new low in perplexity on the datasets of its time and became the first model capable of consistently generating coherent paragraphs; GPT-3 achieved in-context learning; GPT-4, at about 0.91, moved chain-of-thought reasoning from theory into practice. Each decline in the value corresponded to the emergence of a new class of capabilities, and emergence was often sudden rather than linear and gradual—this is precisely the logical foundation for using it as a safety early-warning indicator.

0.68 to 0.7: The Known Unknown

Pachocki offers one of the most forward-looking judgments to date: he expects that deep learning research itself may see “avalanche-like” pretraining progress as early as 2027. More critical is the 0.68 to 0.7 range—OpenAI explicitly states that it expects further “transformative capabilities” to appear in this window, but it does not know what they will be.

A lab that possesses the most advanced technology currently available openly admits that it cannot predict the specific form of the next capability emergence and can only infer from historical patterns that “some major event is about to happen.” Based on this uncertainty, Pachocki proposes the idea of “common limits”: frontier developers would commit that, before entering the 0.68 to 0.7 range, mature mechanisms for monitoring AI behavior—for example, successor technologies to chain-of-thought monitoring—must be in place. In other words, safety capabilities must be in place before the capabilities themselves.

This framework did not emerge out of nowhere. Anthropic CEO Dario Amodei previously proposed a “pace the frontier” plan, calling on companies to adopt common safety standards and to set limits on the pace of unconstrained AI progress. Pachocki’s proposal goes a step further on this basis by providing specific quantitative thresholds and anchoring abstract safety commitments to a measurable number.

The Commercial Cost of Transparency and the Safety Calculus

In the article, Pachocki directly acknowledges that disclosing this set of data “gives competitors a clearer understanding of OpenAI’s position.” His judgment is that the urgency of establishing common industry safety standards outweighs this commercial concern.

From a competitive standpoint, this proposal means very different things for different participants. For OpenAI, proactively disclosing data and proposing industry standards is itself a contest over agenda-setting power—by being the first to define “what constitutes a safe pace,” it may influence future regulatory frameworks. For other frontier labs, accepting this framework means they would likewise need to disclose comparable data, thereby losing the cover of opacity regarding their technical progress; refusing, meanwhile, exposes them to public pressure for being “unwilling to accept safety constraints.” For regulators across the industry, this is the first time a lab has proactively provided a concrete measurement language that could become the technical basis for legislation or administrative rules.

For enterprise users and application-layer developers, this set of figures provides a new reference coordinate system. Which perplexity range a model falls into broadly determines the upper limit of the tasks it is suited to handle: 0.91 (GPT-4 level) is sufficient to support serious reasoning applications; models in the below-0.8 range possess emerging capabilities that are still being explored; around 0.76, models can already produce valuable results on frontier mathematical problems. If this coordinate system gains industry consensus, model selection will move from “intuitive assessment” toward a more evidence-based decision process.

The Metric’s Own Ceiling

Pachocki also lists the limitations of this metric one by one. First, different labs use different test datasets, making direct cross-institutional comparison fundamentally difficult—the figures OpenAI published are based on its own internal held-out datasets and are not inherently comparable with those of other labs. Second, perplexity itself does not correspond to any specific task, and progress in lower perplexity does not necessarily bring capability improvements perceptible to users. Third, capability gains obtained through inference-time compute cannot be reflected in the perplexity metric, meaning the metric will miss part of the progress along the reasoning-model path. Fourth, the lower bound of this metric is unknown; no one knows what numerical value “perfect compression” roughly corresponds to.

These limitations do not negate the value of the proposal itself, but they mean perplexity is better suited as a starting point for safety discussions than as an endpoint. It can tell you “how thick the system’s underlying intelligence substrate is,” but it cannot tell you “how dangerous this capability is in a specific scenario”—the latter still requires task-specific evaluation and deployment-scenario analysis.

Forward-Looking Judgment

Whether this proposal can ultimately form an industry consensus depends on two core variables: first, whether other frontier labs such as Google DeepMind and Anthropic are willing to disclose comparable perplexity data, making cross-institutional comparison possible; second, whether regulators incorporate this framework into governance tools, thereby exerting external pressure on labs that refuse to accept the standards.

If a breakthrough in deep learning research does occur in 2027, whether a common measurement language and pre-established response mechanisms exist at that time may be the key fork between safety discussions moving from “post-mortem review” to “ex ante control.” OpenAI’s choice to proactively disclose the numbers at this point is essentially saying: the clock has started counting down, and everyone may as well use the same watch. Whether this proposal can be implemented, the responses of other labs in the coming months will be the most critical signal.