OpenAI's Largest Pre-Training Model 'Doug' Exposed, Potentially Launching Before November
Claire Weston
OpenAI is building Doug, its largest pre-training run ever, with a launch expected by November — the company's first real attempt at a foundation-model upgrade in nearly two years, rather than stacking reinforcement learning on top of the old base.
What is Doug — and why isn't it GPT-6?
X user ChrisGPT disclosed on August 9 that OpenAI is advancing a new pre-training model codenamed Doug — described as OpenAI's largest pre-training effort to date.
Doug and GPT-6 are not the same model. ChrisGPT said GPT-6 is likely Astra, the internal model whose release was paused over safety concerns.
This means → OpenAI may be running at least two major model tracks in parallel: Astra, already in advanced evaluation, and the even larger Doug.
Why does OpenAI need a new large-scale pre-training run?
After GPT-4o shipped in May 2024, OpenAI went nearly two years without completing a full-scale pre-training run deployable as a new flagship.
Capability gains in o1, o3, and the GPT-5 series came mainly from large-scale reinforcement learning and inference-time compute — giving the model more processing power to "think" before answering.
In plain terms = for two years OpenAI has been adding floors to an old foundation. RL pushed the old base to its limits, but the foundation itself never changed. Doug's job is to lay a new one.
Where does Doug's technical lineage come from?
In December 2025, *The Information* reported that OpenAI had developed a pre-training model codenamed Garlic, which performed well on coding and reasoning benchmarks and incorporated fixes for problems discovered in earlier training runs.
Chief Research Officer Mark Chen reportedly told his team that several key pre-training issues had been resolved.
In January 2026, research firm SemiAnalysis independently confirmed the progress. In plain terms = Garlic was the test plot — it proved the fixes worked. Doug is what happens when those methods are scaled up.
What role did SemiAnalysis play?
As early as July 9, SemiAnalysis wrote in a memo to institutional clients that OpenAI had overcome its pre-training problems and that a much larger model codenamed Doug was actively under way.
That memo went public on August 7 alongside an article on Gemini and Google Cloud — two days before ChrisGPT's disclosure.
This reflects a longer information trail: industry research had tracked Doug for at least a month before the public reveal.
What is the real question Doug has to answer?
The past two years proved that the old foundation can still deliver capability gains through reinforcement learning and inference-time compute.
Doug addresses the next-order question: once the foundation itself takes a major leap, how much further can the post-training stack — already pushed to its limits — carry performance?
This means → Doug's outcome will not just determine one model's quality. It will set OpenAI's true starting line for the next round of model competition — whether the new foundation can support a taller building.
Content is for reference only, not financial advice.