OpenAI Doug Leak: Largest Pre-Training Model in Development
OpenAI is reportedly advancing a new large-scale foundation model codenamed “Doug”, which could become the company’s largest pre-training effort to date. According to industry reports and model-development trackers, Doug is separate from the rumored GPT-6 successor codenamed “Astra” and represents a renewed push toward scaling base model training.
If the reports are accurate, Doug would mark a strategic shift for OpenAI after nearly two years of emphasizing post-training, reinforcement learning (RL), and inference-time compute optimization rather than launching a completely new large-scale pre-trained foundation model.
The project highlights a broader question facing the AI industry: after extracting significant gains from reasoning models and inference scaling, can another major leap in capability come from returning to massive foundation model pre-training?
🧠The Emergence of OpenAI’s Doug Model #
On August 9, 2026, AI model tracker ChrisGPT reported that OpenAI is actively developing a large pre-training model internally known as Doug.
The reported details suggest:
- Doug is separate from GPT-6.
- GPT-6 is believed to correspond to Astra, a flagship model undergoing additional safety and capability evaluation.
- Doug represents a much larger-scale pre-training initiative.
- The model could arrive as early as November 2026.
The codename has also appeared in third-party industry analysis. SemiAnalysis reportedly referenced an internal research memo discussing OpenAI’s pre-training progress and stated that OpenAI had overcome previous pre-training challenges while developing a significantly larger model under the Doug codename.
If validated, Doug would represent a major milestone because it indicates OpenAI may have resolved technical and infrastructure challenges that previously slowed large-scale foundation model development.
🔄 OpenAI’s Shift Toward Reinforcement Learning and Inference Scaling #
OpenAI’s recent model strategy has focused heavily on improving reasoning capabilities through post-training methods rather than relying only on larger pre-trained models.
The transition began prominently with GPT-4o and continued through reasoning-focused model families.
GPT-4o as the previous foundation #
Released in May 2024, GPT-4o became OpenAI’s primary flagship model and served as the foundation for several subsequent improvements.
Instead of immediately replacing it with a completely new large-scale pre-trained model, OpenAI increasingly explored alternative scaling methods.
These included:
- Reinforcement learning-based reasoning optimization.
- Increased inference-time computation.
- Specialized reasoning models.
- Dynamic model routing architectures.
o1 introduced reasoning-time scaling #
In September 2024, OpenAI introduced o1-preview, demonstrating a new approach where models could use additional computation during inference to improve reasoning performance.
Rather than only increasing parameter counts and training data, the strategy focused on teaching models how to allocate more computational effort when solving difficult problems.
o3 expanded reinforcement learning approaches #
In April 2025, OpenAI continued expanding this direction with o3, reinforcing the importance of large-scale reinforcement learning for advanced reasoning.
The approach showed that post-training optimization could unlock significant improvements without immediately requiring an entirely new foundation model.
GPT-5 unified multiple AI capabilities #
In August 2025, OpenAI introduced GPT-5 with a unified architecture combining:
- Fast-response models.
- Deep reasoning engines.
- Intelligent routing systems.
This represented the culmination of the inference and post-training scaling strategy.
However, according to industry analysis, these models were reportedly built on improvements derived from existing foundations rather than a completely new generation of base model pre-training.
🧄 Project Garlic Helped Resolve Pre-Training Challenges #
The reported development path toward Doug reportedly began with another internal project known as Garlic.
Following Google’s release of Gemini 3 in November 2025, OpenAI reportedly initiated a company-wide response focused on improving ChatGPT performance and accelerating model development.
Garlic as a validation platform #
According to reports, Garlic served as a smaller-scale pre-training effort used to address technical bottlenecks.
The project reportedly delivered improvements in areas including:
- Coding performance.
- Reasoning benchmarks.
- Training stability.
- Knowledge efficiency.
One reported breakthrough was that OpenAI researchers found methods allowing smaller models to absorb capabilities that previously required significantly larger architectures.
This type of improvement is particularly valuable because it can reduce training inefficiency while enabling future models to scale more effectively.
From Garlic to Doug #
Industry analysis suggested that Garlic was not intended to be the final frontier model. Instead, it functioned as a validation stage that demonstrated OpenAI’s ability to overcome previous pre-training limitations.
Doug would reportedly represent the next step: applying these improvements to a much larger foundation model.
📈 OpenAI Returns to Large-Scale Foundation Model Scaling #
If current reports are accurate, OpenAI is now operating multiple major model development pipelines:
- Astra: A flagship capability model undergoing safety and deployment evaluation.
- Doug: A large-scale foundation model focused on renewed pre-training expansion.
This represents a return to a traditional AI scaling strategy where improvements come from increasing the size, quality, and capability of the underlying model itself.
Why pre-training still matters #
Post-training and reinforcement learning have demonstrated that existing models can become significantly more capable through improved reasoning strategies and additional computation.
However, foundation model limitations eventually become a bottleneck.
A stronger base model can provide improvements across:
- General knowledge.
- Coding ability.
- Scientific reasoning.
- Multimodal understanding.
- Long-context performance.
- Adaptability through post-training.
The combination of a substantially improved base model with advanced reinforcement learning techniques could unlock another major capability jump.
🚀 Doug Could Define OpenAI’s Next Scaling Era #
The reported Doug project represents a possible transition point in OpenAI’s model strategy.
The company has spent recent years proving that reasoning optimization and inference-time scaling can dramatically improve AI performance. Doug would test a different hypothesis: whether returning to aggressive foundation model pre-training can create another major capability breakthrough.
If OpenAI successfully combines a larger next-generation pre-trained model with its existing reinforcement learning infrastructure, reasoning systems, and deployment ecosystem, Doug could become the foundation for the company’s next AI platform generation.
For the broader AI industry, the project also signals that large-scale pre-training remains a central competitive battlefield. Even as reasoning models and agent systems gain attention, the underlying foundation model continues to determine the upper limits of future AI capabilities.