OpenAI is said to be pushing forward with a new large-scale pre-training model codenamed Doug, according to a post published on Aug. 9 by X user ChrisGPT, who has closely tracked the company’s model roadmap.
ChrisGPT said Doug would be OpenAI’s biggest pre-training project so far and not the same model as GPT-6. In follow-up replies, he added that GPT-6 is likely Astra, the company’s strongest in-house model whose release was reportedly paused a day earlier over safety concerns. He also said Doug could arrive by November at the latest.
Doug had already appeared in public reporting
ChrisGPT was not the first public source to mention Doug.
On Aug. 7, semiconductor and AI research firm SemiAnalysis published an article about Gemini and Google Cloud that included an excerpt from a research memo previously sent to institutional clients. The memo was dated July 9. The article is available at https://newsletter.semianalysis.com/p/gemini-is-cooked-but-gcp-is-cooking .
One line stood out: OpenAI had overcome its pre-training problems, and a much larger model codenamed Doug was being actively worked on.
If that information is accurate, Doug would suggest OpenAI is reviving a large-scale base-model transition after nearly two years in which capability gains were driven mainly by post-training, reinforcement learning, and inference-time compute.
From GPT-4o onward, more gains were attributed to RL
The timeline in the report starts with GPT-4o.
On May 13, 2024, OpenAI released GPT-4o and called it its new flagship model. Over roughly the next two years, OpenAI trained and released newer pre-training models including GPT-4.5, but it did not complete a full-scale pre-training run that was widely deployed as the next main frontier model.
During that period, visible progress in OpenAI’s systems increasingly came from another route.
On Sept. 12, 2024, OpenAI released o1-preview. Rather than relying mainly on bigger pre-training runs, o1 showed another scaling path: large-scale reinforcement learning that taught the model to spend more compute on reasoning. After that, post-training, RL, and inference-time compute became more central across OpenAI’s model stack.
In April 2025, OpenAI formally released o3 and again said the reasoning performance of the o-series came from large-scale reinforcement learning.
In August 2025, GPT-5 was released. It was described not as a single model but as a unified architecture made up of a fast model, a deep reasoning model, and a routing system.
SemiAnalysis argued that o1, o3, and even the GPT-5 family did not come with a new base-model generation comparable to GPT-4o. In its view, those systems were still built on the foundation established in the GPT-4o era.
OpenAI has never confirmed that training lineage. Still, if SemiAnalysis is correct, the company’s path over the past two years becomes easier to read: the underlying base model did not undergo a comparable generational jump, while capability gains came mostly from stronger post-training and RL.
That would also mean model scaling expanded from a framework centered mainly on pre-training to one spanning three dimensions: pre-training, RL, and inference-time compute.
The report says the risk in that approach is diminishing returns if the base model does not get a similar upgrade for an extended period and the company keeps leaning on post-training and inference compute alone.
Garlic was presented as a validation stage for pre-training fixes
The arrival of Gemini 3 was described in the article as the point where that training-route issue turned into direct competitive pressure.
On Nov. 18, 2025, Google released Gemini 3. Ten days later, SemiAnalysis wrote in its TPUv7 analysis that OpenAI had not completed, since GPT-4o, a successful full-scale pre-training run that could be widely deployed as a new frontier model. Once Google launched Gemini 3, the gap was no longer just about training direction. It became a competitive issue.
On Dec. 1, multiple media outlets reported that Sam Altman had declared a "Code Red" internally at OpenAI, asking teams to prioritize ChatGPT and reallocate some resources.
A day later, more training details surfaced.
On Dec. 2, 2025, The Information reported that OpenAI was developing a new pre-training model codenamed Garlic. Citing internal sources, the report said Garlic was performing well in coding and reasoning evaluations and used a series of bug fixes discovered during earlier training efforts.
The report also said OpenAI Chief Research Officer Mark Chen had told the team that the company had solved some key problems in pre-training. Those improvements, according to the report, made it possible for smaller models to absorb knowledge that previously required larger models.
The same report included another line that later took on more weight: OpenAI had already started developing an "even bigger and better model" based on what it learned from Garlic. In this account, that is where the Doug story really begins.
SemiAnalysis repeated the point in January 2026
On Jan. 6, 2026, SemiAnalysis returned to OpenAI’s model strategy and wrote directly that the company had solved its pre-training issues.
Under that reading, the obstacles that had hampered OpenAI’s full-scale pre-training efforts had been removed. The article argues that Garlic likely served to test whether those fixes worked, while Doug may be the result of scaling those methods to a larger training run.
Doug points to another round of base-model scaling
If the reports are right, OpenAI may now be advancing at least two major model programs at the same time: Astra, which the article describes as already in an advanced evaluation stage, and Doug, which is said to be even larger.
Doug matters for another reason. It would indicate that OpenAI is preparing to push scaling at the foundation-model layer again.
Over the past two years, OpenAI has shown that an older base can still be pushed upward with RL, reasoning systems, and inference-time compute. The question attached to Doug is different: if the base itself makes another large jump, how far can that heavily optimized post-training system take model capability next?
That, in the article’s framing, could mark the starting point of OpenAI’s next round of model competition.
References cited in the source article
- ChrisGPT on X: https://x.com/ChrisGPT/status/2086220662264250764
- SemiAnalysis article: https://newsletter.semianalysis.com/p/rl-environments-and-rl-for-science
- The Information report: https://www.theinformation.com/newsletters/ai-agenda/openai-developing-garlic-model-counter-googles-recent-gains
The source article said the piece originally came from the WeChat public account "机器之心" with the ID almosthuman2014, and was authored by "关注LLM的机器之心."