OpenAI appears to be restarting large-scale foundational model development after nearly two years of focusing on post-training and reasoning improvements, according to reports about a project codenamed Doug that could launch by November. This shift represents a significant strategic pivot for the company, which has spent the past 24 months proving that older base models could be continuously improved through reinforcement learning (RL), reasoning, and inference-time compute rather than building entirely new foundational models from scratch.

What Is Doug and Why Does It Matter?

Doug is described as OpenAI's largest pre-training project to date, distinct from other models in development like Astra and GPT-6. The project was first publicly mentioned by ChrisGPT, an X user who tracks OpenAI developments, on August 9, with expectations for a launch no later than November. SemiAnalysis, a semiconductor and artificial intelligence research firm, had previously reported in July that OpenAI had overcome pre-training challenges and was actively advancing a much larger model codenamed Doug.

The significance of Doug lies in what it represents strategically. Since May 2024, when OpenAI released GPT-4o as its flagship model, the company has not completed a full-scale pre-training round that could be widely deployed as a next-generation frontier model. Instead, capability improvements came from three other dimensions: post-training refinements, reinforcement learning systems, and inference-time compute, which allows models to spend more computational resources thinking through problems at the moment of use rather than during initial training.

How Did OpenAI Solve Its Pre-Training Problem?

According to reports from The Information in December 2025, OpenAI was developing a model codenamed Garlic that performed well in coding and reasoning evaluations while incorporating bug fixes discovered during previous training attempts. Mark Chen, OpenAI's Chief Research Officer, reportedly told the team that the company had resolved key issues in previous pre-training efforts. These improvements allegedly allowed smaller models to hold knowledge that previously required much larger models to acquire.

By January 2026, SemiAnalysis reported that OpenAI had resolved the pre-training issues that had previously plagued full-scale model development. Garlic likely served as a verification project for these fixes, while Doug represents the scaling up of these corrected training methods to a much larger scale. This progression suggests OpenAI has moved from identifying problems to validating solutions to implementing them at production scale.

Steps to Understanding OpenAI's Model Development Strategy

• Pre-Training Phase: The foundational step where models learn from massive amounts of raw text data before any specialized fine-tuning, which OpenAI had struggled with but claims to have resolved through Garlic and related projects.
• Post-Training Refinement: After pre-training, models undergo reinforcement learning and other optimization techniques to improve reasoning, safety, and task-specific performance, the approach OpenAI relied on for the past two years.
• Inference-Time Compute: Allowing models to allocate more computational resources to thinking through complex problems at the moment of use, rather than baking all reasoning into the training process itself.

Why Is Base Model Scaling Becoming Urgent Again?

The competitive landscape shifted dramatically with Google's release of Gemini 3, which turned OpenAI's training roadmap challenges into direct competitive pressure. In December 2025, Sam Altman reportedly declared "Code Red" internally at OpenAI, requiring teams to prioritize improving ChatGPT and reallocating resources accordingly. This urgency reflects a fundamental concern: if foundational models do not undergo generational upgrades and companies rely solely on post-training and inference compute to continue scaling, they will eventually face diminishing marginal returns.

Doug answers a critical question for OpenAI's next phase of competition: after the base model itself completes a major leap again, how far can the post-training systems, which have been pushed to their limits, take model capabilities ? This may represent the real starting point for OpenAI's next round of model competition against rivals like Google and others developing advanced reasoning systems.

The emergence of Doug also suggests OpenAI is pursuing at least two significant model projects simultaneously. Astra, which has entered advanced evaluation stages, represents one track, while Doug represents a parallel effort focused on restarting foundational model scaling. This dual approach reflects the company's assessment that both immediate capability improvements and longer-term foundational advances are necessary to maintain competitive advantage in the rapidly evolving artificial intelligence landscape.