August 9th: X user ChrisGPT, who has long been tracking OpenAI's model developments, leaked that OpenAI is advancing a new large-scale pre-trained model codenamed Doug.
According to his claims, Doug will be OpenAI's largest pre-training project to date, and it is not the same model as GPT-6.
In subsequent replies, ChrisGPT further indicated that GPT-6 is very likely Astra, OpenAI's most powerful model whose release was urgently paused yesterday due to safety concerns.
As for Doug, he expects it could be released no later than November.
ChrisGPT is not the first source to publicly mention Doug.
On August 7th, semiconductor and AI research firm SemiAnalysis, in an article discussing Gemini and Google Cloud, publicly disclosed a segment of a research memorandum previously sent to its institutional clients. This memo was dated July 9th.
Article address: https://newsletter.semianalysis.com/p/gemini-is-cooked-but-gcp-is-cooking
There is one crucial sentence: OpenAI has overcome its pre-training issues, and a model codenamed Doug, which is much larger, is being actively advanced.
If the related information is accurate, Doug might signify that, after nearly two years of relying primarily on post-training, reinforcement learning, and inference-time compute to drive capability growth, OpenAI is restarting a large-scale base model generational upgrade.
The story starts with GPT-4o.
OpenAI Shifted More Growth to RL
On May 13, 2024, OpenAI released GPT-4o, calling it its new flagship model.
For nearly two years since then, although OpenAI has trained and released new pre-trained models like GPT-4.5, it never completed a full-scale pre-training round that could be widely deployed as the next-generation main frontier model.
Meanwhile, OpenAI's model capability growth increasingly came from another route.
On September 12, 2024, OpenAI released o1-preview.
Compared to the past reliance on larger-scale pre-training to drive capability growth, o1 demonstrated another scaling method: through large-scale reinforcement learning, let the model learn to invest more compute in reasoning. Subsequently, post-training, RL, and inference-time compute became increasingly important in OpenAI's model system.
In April 2025, o3 was officially released. OpenAI emphasized again that the reasoning capabilities of the o-series come from large-scale reinforcement learning.
In August 2025, GPT-5 was released. It was no longer just a single model but a unified architecture consisting of a fast model, a deep reasoning model, and a routing system.
SemiAnalysis believes that behind the o1, o3, and even the GPT-5 series, there was no accompanying full-scale new base model generational leap equivalent to GPT-4o. The related models were actually still built upon the base model system from the GPT-4o era.
OpenAI never confirmed this training lineage, but if SemiAnalysis's information holds, the strategy of the past nearly two years becomes easy to understand: the foundation did not undergo a generational leap of the same magnitude; capability growth was primarily driven by increasingly stronger post-training and RL.
Model scaling also gradually expanded from relying mainly on pre-training in the past to three dimensions: pre-training, RL, and inference-time compute.
The problem is, if the base model does not undergo a generational upgrade of the same magnitude for a long time, relying only on post-training and inference compute to continue scaling will eventually face diminishing marginal returns.
And the emergence of Gemini 3 rapidly transformed this potential training strategy issue into real competitive pressure.
Garlic: Pre-training Starts Running Again
On November 18, 2025, Google released Gemini 3.
Ten days later, SemiAnalysis threw out a widely discussed assessment in their TPUv7 analysis: Since GPT-4o, OpenAI has not completed a single successful full-scale pre-training that could be widely deployed as a new frontier model.
When Google launched Gemini 3, this difference also shifted from a training strategy problem to direct competitive pressure.
On December 1st, multiple media outlets reported that Sam Altman internally announced a "Code Red" at OpenAI, requiring teams to prioritize improving ChatGPT and reallocating some resources.
A day later, more critical training information surfaced.
On December 2, 2025, *The Information* reported that OpenAI was developing a new pre-trained model codenamed Garlic. The report cited internal sources stating that Garlic performed well on coding and reasoning benchmarks while incorporating a series of bug fixes discovered by OpenAI during previous training runs.
More crucially, OpenAI Chief Research Officer Mark Chen reportedly told the team that the company had resolved some key issues in previous pre-training. The report also mentioned that these training improvements could allow smaller models to hold knowledge that previously required larger models.
In the same article, there was another sentence that later seemed particularly important: OpenAI has already begun developing an "even bigger and better model" based on the experience learned from Garlic.
The story of Doug essentially started here.
On January 6, 2026, SemiAnalysis again discussed OpenAI's model roadmap. This time, they directly wrote: OpenAI has solved its pre-training issues.
In other words, according to the information SemiAnalysis possesses, the issues that previously plagued OpenAI's full-scale pre-training have been resolved.
Garlic likely served the role of verifying whether these fixes were effective, and Doug could be the result after these training methods were truly scaled up to a larger size.
OpenAI May Be Preparing to Restart Base Scaling
If the above information is accurate, then OpenAI might be advancing at least two significant model projects in succession: Astra, which has entered advanced evaluation stages, and Doug, which is reportedly larger in scale.
And Doug points to another thing: restarting scaling of the base model itself.
Over the past two years, OpenAI has proven that an old base can still be pushed upward through RL, reasoning, and inference-time compute.
Doug aims to answer another question: After the base itself undergoes another major leap, how far can this already-pushed-to-the-limit post-training system take its capabilities.
This might be the real starting point for OpenAI's next round of model competition.
References:
https://x.com/ChrisGPT/status/2086220662264250764
https://newsletter.semianalysis.com/p/gemini-is-cooked-but-gcp-is-cooking
https://newsletter.semianalysis.com/p/rl-environments-and-rl-for-science
https://www.theinformation.com/newsletters/ai-agenda/openai-developing-garlic-model-counter-googles-recent-gains
This article is from the WeChat public account "Almost Human" (ID: almosthuman2014), author: Almost Human who follows LLMs.
