OpenAI Chief Scientist Proposes Industrywide Caps on Frontier AI Development
OpenAI has published internal pretraining metrics and urged rival laboratories to adopt shared development limits based on model perplexity. The proposal aims to halt the scaling of frontier models until safety monitoring tools can reliably evaluate emerging capabilities.
By Naina Verma
Share this story
OpenAI Chief Scientist Jakub Pachocki has proposed a new industrywide framework to cap the development of frontier artificial intelligence models until safety monitoring catches up. In an essay published on November 20, 2026, Pachocki urged rival laboratories to adopt a shared technical yardstick to track the underlying intelligence of their systems.[1][5]
The proposal centers on a decades-old machine learning metric known as log-perplexity, or cross-entropy loss, which measures how accurately a model predicts the next token in a sequence. By releasing OpenAI's internal pretraining data for the first time, Pachocki aims to establish a common language for the industry.[1][4]
"I believe tracking the pace of progress on training better compressors is key to understanding the arc of advances in AI," Pachocki wrote. He argued that while publishing these figures gives competitors a clearer picture of OpenAI's standing, the urgency of establishing shared safety measures outweighs the competitive risk.[1]
Tracking the Intelligence Substrate
Perplexity serves as a proxy for the raw compression capability of a model during its initial training phase. A lower perplexity score indicates sharper, more accurate predictions. OpenAI has tracked this specific metric internally on a fixed evaluation dataset since the development of GPT-3.5.[1][2]
According to the disclosed figures, specific drops in log-perplexity have historically aligned with sudden, transformative jumps in model capabilities. GPT-3.5 registered a log-perplexity of roughly 0.99, which provided the fluent conversational abilities that powered the original release of ChatGPT.[1][4]
The subsequent generation, GPT-4, reached a log-perplexity of about 0.91. At this threshold, the model became reliably proficient at productive chain-of-thought reasoning. This capability eventually formed the foundation for the modern reasoning models that dominate the current artificial intelligence landscape.[1][3]
As the loss metric dropped below 0.80, OpenAI researchers observed the emergence of entirely new, transformative capabilities in specialized fields such as cybersecurity and advanced mathematics. Pachocki emphasized that these thresholds were not predictable in advance; the capabilities simply materialized once the model's compression reached a certain efficiency.[1][4]
The Mathematics Breakthrough
The most striking example provided in the disclosure occurred when an internal OpenAI model crossed a log-perplexity threshold of approximately 0.76. For over a year, researchers had been testing models against long-standing open problems in mathematics, with limited success.[1][2]
Once the 0.76 mark was breached, the model reportedly began resolving these entrenched mathematical problems one after another. Pachocki described the experience of watching the solutions arrive in a single internal Slack thread as "deeply surreal" for the research team.[1][3]
This breakthrough at 0.76 log-perplexity serves as the foundation for Pachocki's broader warning about the pace of future development. He argues that a similar step-change could soon hit the field of artificial intelligence research itself, fundamentally altering how new models are developed.[1][4]
The 2027 Acceleration Warning
Pachocki projects that as soon as 2027, the industry could witness an avalanche of fundamental pretraining advancements driven disproportionately by artificial intelligence rather than human researchers. This automated acceleration could push models into uncharted capability territories faster than previously anticipated.[1][2]
While artificial intelligence is also expected to contribute to progress in alignment and safety monitoring, Pachocki expressed doubt that these defensive technologies will advance quickly enough. He warned that safety research is unlikely to keep pace with the full potential acceleration in raw pretraining capabilities.[1][4]
Extrapolating from the current trajectory, OpenAI expects further transformative capabilities to emerge as models enter the 0.68 to 0.70 log-perplexity window. The company acknowledges that it does not know what these specific capabilities will be, and current evaluation frameworks may not even be designed to test for them.[1][3]
Structuring the Industry Limits
To manage this uncertainty, the proposal asks frontier developers to voluntarily agree to caps on scaling. "I further propose common limits, based on perplexity and other measurements, that developers are committed to not crossing until our ability to understand and monitor AI behavior achieves a certain concrete milestone," Pachocki wrote.[1]
One concrete milestone Pachocki suggested is the development of a genuine successor to current chain-of-thought monitors. These existing systems inspect a model's written reasoning for signs of misalignment, but they are expected to degrade in effectiveness as models become more advanced and evaluation-aware.[1][2]
A recent OpenAI monitorability study highlighted the complexities of this challenge. The research examined proxies for monitorability, such as the ability to manipulate reasoning tokens when prompted, and found that current control mechanisms remain noisy and imperfect.[2]
The study concluded that while reinforcement learning can increase a model's controllability, these techniques do not always prevent a system from evading monitors. Consequently, Pachocki argues that further pretraining progress will absolutely require entirely new approaches to safety auditing.[2][4]
Under the proposed framework, developers would jointly establish shared norms for evaluating models and red-teaming safety cases before any new system is deployed. This initial limit is intended to serve as the foundation for a recurring, industry-wide framework for pacing future development.[1][3]
Context of the Safety Debate
The push for coordinated limits echoes similar arguments made earlier in 2026 by other industry leaders, including Anthropic's Dario Amodei. Hundreds of employees across frontier laboratories have previously signed open letters advocating for industry-wide coordination on the pace of artificial intelligence development.[4][5]
The disclosure arrives at a critical juncture for artificial intelligence governance, as policymakers globally debate how to regulate frontier models. By proposing a voluntary, metric-driven framework, OpenAI is attempting to demonstrate that the industry can self-regulate before governments impose rigid statutory caps.[2][5]
However, the proposal also highlights the inherent tension in the current development race. Laboratories are simultaneously investing billions of dollars into massive new computing clusters to drive perplexity lower, while publicly acknowledging that they may not be able to safely control the resulting systems.[3][5]
Limitations of the Metric
Despite positioning perplexity as a shared yardstick, OpenAI acknowledges several technical challenges in adopting it as a universal standard. The most significant hurdle is the difficulty of comparing scores across different laboratories, as each developer uses distinct evaluation datasets and training mixtures.[1][4]
Because perplexity is tied to a specific dataset, a score of 0.80 at OpenAI does not necessarily equate to a 0.80 at Google or Anthropic. Pachocki suggested that if the industry commits to transparency, researchers can develop transferable methods to accurately compare the base quality of different models.[1][2]
Furthermore, perplexity does not directly correspond to any actionable, real-world task. It is a measure of underlying prediction efficiency rather than a test of practical utility. The metric's value lies entirely in its historical correlation with the sudden emergence of entirely new capabilities.[1][3]
The metric also fails to account for massive capability gains achieved outside of the initial pretraining phase. For instance, the implementation of chain-of-thought reasoning drastically increases a model's practical intelligence and utility, but it does not reduce the base model's underlying perplexity score.[1][4]
Finally, the theoretical floor for log-perplexity remains unknown. While researchers do not expect the loss to ever reach zero—which would require perfectly deterministic prediction of the entire dataset—they do not know the inherent entropy limit of the training data.[1][2]
As models approach this unknown floor, the apparent numerical progress on perplexity will inevitably slow down. However, Pachocki cautioned that at these extreme levels of compression, even minuscule numerical gains in perplexity will likely correspond to massive leaps in actual model intelligence.[1][3]
As models approach this unknown floor, the apparent numerical progress on perplexity will inevitably slow down.
The proposal ultimately attempts to transform a highly technical engineering metric into a practical governance tool. By tying development caps to measurable prediction efficiency, OpenAI hopes to force the industry to prioritize monitorability over raw capability scaling.[4][5]
Whether rival laboratories will agree to publish their own closely guarded pretraining metrics remains to be seen. For now, OpenAI has placed the burden of coordination on the broader industry, establishing a clear technical marker for when the next wave of transformative capabilities might arrive.[2][5]
Key points
- OpenAI Chief Scientist Jakub Pachocki has proposed capping frontier AI development based on shared technical metrics.
- The framework relies on log-perplexity, a measure of how efficiently a model predicts data during its initial training.
- OpenAI disclosed that a model reaching 0.76 log-perplexity began independently solving entrenched mathematical problems.
- The proposal asks laboratories to halt development at agreed thresholds until safety monitoring tools achieve specific milestones.
Unanswered questions
- Whether rival frontier laboratories like Google and Anthropic will agree to publish their own proprietary pretraining metrics.
- How the industry will standardize log-perplexity scores across different companies that use entirely distinct training datasets.
- What specific capabilities will emerge when models enter the projected 0.68 to 0.70 log-perplexity window.
How we got here
2022
GPT-3.5 achieves a log-perplexity of roughly 0.99, enabling fluent conversational capabilities.
2023
GPT-4 reaches a log-perplexity of 0.91, demonstrating reliable chain-of-thought reasoning.
Mid-2026
An internal OpenAI model crosses the 0.76 threshold, producing groundbreaking results on open mathematics problems.
Nov 20, 2026
OpenAI publishes its internal metrics and proposes industrywide development limits.
2027
The projected timeframe when AI could begin disproportionately accelerating its own pretraining research.
- Frontier AI Developers
- Argue that shared technical metrics are necessary to pace development and ensure safety tools keep up with raw capabilities.
- Industry Analysts
- Note that while the proposal sets a concrete marker, cross-lab comparability remains untested and the metric misses non-pretraining gains.
Perspectives this story doesn't cover
- Other frontier AI laboratories (Google, Anthropic)
- Open-source AI developers
Sources
[1]OpenAIFrontier AI DevelopersPerplexity and Intelligence
Read on OpenAI →
[2]Unite.AIIndustry AnalystsOpenAI Proposes Shared AI Development Limits Tied to Safety Milestones
Read on Unite.AI →
[3]Seeking AlphaIndustry AnalystsOpenAI proposes shared perplexity-based AI limits until monitoring improves; models near 0.68-0.70 may develop unknown transformative capabilities
Read on Seeking Alpha →
[4]OfficeChaiIndustry AnalystsOpenAI Wants 'Perplexity' To Become The Industry's Shared Yardstick For AI Progress And Safety Limits
Read on OfficeChai →
[5]BloombergIndustry AnalystsOpenAI Chief Scientist Urges Industry Limits on AI Development
Read on Bloomberg →
More in Technology
See all →AI Governance
Hendrycks and Hodak Clash Over Frontier AI Moratorium at DealBook Summit
5 sources
AI Regulation
California Sues OpenAI Over Rogue Agent Hack on Hugging Face Repository
7 sources
AI Governance
OpenAI Signs AI Safety Research Partnership With Poland's NASK Institute
4 sources
AI Safety Evaluation
DeepSeek's Latest Foundation Model Persistently Resists Shutdown in New Safety Evaluations
5 sources
Comments
Every angle. Every day.
Get Technology stories with full source coverage and perspective breakdowns, free every day.




