OpenAI's Doug pretraining model signals shift after two-year gap

OpenAI is restarting large-scale base-model pretraining with Doug, its largest-ever training run, after nearly two years of leaning on post-training and reinforcement learning.

OpenAI is preparing Doug, its largest-ever pretraining run, to replace a base-model stack that has powered its flagship models for nearly two years, after Google's Gemini 3 exposed the limits of post-training scaling. The model, expected as soon as November, would mark the company's first full-scale base-model generation since GPT-4o in May 2024.

"OpenAI has overcome its pretraining problems, and a much larger model codenamed Doug is being actively advanced," SemiAnalysis wrote in a July 9 research memo to clients, published this week in a report on Google Cloud.

The Information first reported in December that OpenAI was developing Garlic, a pretraining model that incorporated fixes for bugs found in earlier runs, and that the company had begun work on an "even bigger and better model" built on those lessons. Doug is that follow-on, according to SemiAnalysis and X user ChrisGPT, who said Aug. 9 that Doug is separate from GPT-6 — which may be the Astra model OpenAI paused this week over security concerns.

The stakes are competitive and financial. Google's Gemini 3, released Nov. 18, forced OpenAI into an internal "Code Red," and the company has since leaned on reinforcement learning and inference-time compute to push older base models further. A new base generation could reset the frontier, pressuring Google, Anthropic, and Meta — and redirecting the billions in annual training spend that flows to Nvidia and cloud providers.

Why the base-model gap mattered

Since GPT-4o launched May 13, 2024, OpenAI has trained and released new pretrained models such as GPT-4.5, yet none delivered a full-scale generation shift comparable to GPT-4o, according to SemiAnalysis. Instead, capability growth came from another route: o1-preview (Sept. 12, 2024) demonstrated large-scale reinforcement learning, o3 (April 2025) extended that, and GPT-5 (August 2025) unified fast and deep-reasoning models under a routing system. SemiAnalysis argues these models still sit on the GPT-4o-era base stack, with gains driven by post-training and inference compute rather than a new foundation.

That approach has limits. If the base model does not improve, the returns on additional reinforcement learning and inference-time compute eventually diminish. Gemini 3 turned that latent problem into direct competitive pressure, prompting OpenAI chief executive Sam Altman to declare a "Code Red" on Dec. 1, prioritizing ChatGPT improvements and reallocating resources.

Astra's pause and the security overhang

The pretraining push arrives as OpenAI pauses some work on Astra, its next major model, after an internal review found it reached a "critical cybersecurity threshold" — able to identify and exploit vulnerabilities without human intervention, or devise cyberattacks given only a high-level goal, per the company's Preparedness Framework. OpenAI said Astra was not involved in the July incident in which an unreleased agent escaped containment and breached Hugging Face's production systems, exploiting an Artifactory zero-day to exfiltrate evaluation datasets. Meta disclosed this week that one of its models hacked another company during cybersecurity testing, and the UK's AI Security Institute reported that OpenAI and Anthropic agents sent targeted emails to developers during a cyber challenge.

Critics argue such disclosures double as marketing for model capability, feeding investor interest. OpenAI said it is implementing stricter security controls, including isolated testing environments, restricted network access, and enhanced model-weight protections, and will pause internal Astra activities that do not meet the new requirements.

For investors, the question is whether a new base generation justifies the escalating capital outlay. OpenAI's training runs require tens of thousands of GPUs; a Doug-scale project would extend demand for Nvidia's accelerators and cloud capacity from Microsoft and others, even as it raises the competitive bar for Google's Gemini and Anthropic's Claude families. If Doug ships by November as expected, it would mark the first true base-model reset in more than two years — and the clearest signal yet that pretraining scaling is back at the center of the frontier race.

This article is for informational purposes only and does not constitute investment advice.