OpenAI Unveils Massive Pre‑training Model ‘Doug’ – Is a New Base Model Finally Arriving?

On August 9, a user known as ChrisGPT on X reported that OpenAI is advancing a new large‑scale pre‑training model codenamed Doug , which is claimed to be the biggest pre‑training effort to date and distinct from the rumored GPT‑6 model.

Earlier, the research firm SemiAnalysis had circulated a memo (dated July 9) stating that OpenAI had overcome pre‑training challenges and was actively pushing a much larger model named Doug. The memo’s key line reads: “OpenAI has overcome pre‑training issues, a much larger model codenamed Doug is in active development.”

The analysis traces the narrative back to GPT‑4o, released on May 13 2024 as OpenAI’s flagship model. Over the subsequent two years, OpenAI released incremental models such as GPT‑4.5 but did not complete a full‑scale pre‑training round that could serve as a new frontier model. Instead, capability gains shifted toward post‑training, reinforcement learning (RL), and inference‑time compute.

Evidence of this shift includes the release of o1‑preview on September 12 2024, which demonstrated scaling via large‑scale RL that forces the model to allocate more computation during inference. Subsequent models o3 (April 2025) and GPT‑5 (August 2025) continued this trend, with GPT‑5 comprising a suite of fast, deep‑reasoning, and routing components rather than a single monolithic base.

SemiAnalysis argues that none of these later models represent a full‑scale base‑model jump comparable to GPT‑4o; they build on the same underlying architecture while augmenting it with stronger post‑training and RL techniques. The firm warns that relying solely on these methods without a fresh base may eventually encounter diminishing returns.

Competitive pressure intensified when Google announced Gemini 3 on November 18 2025. SemiAnalysis later highlighted that, since GPT‑4o, OpenAI has not delivered a widely deployable new frontier model, a gap that Gemini 3 begins to exploit.

Following internal reports of a “Code Red” on December 1 2024—where Sam Altman ordered a priority boost for ChatGPT—further leaks emerged. The Information reported on December 2 2025 that OpenAI is developing a new pre‑training model called Garlic**, which shows strong performance on coding and reasoning benchmarks and incorporates bug fixes discovered in earlier training runs.

OpenAI’s chief research officer Mark Chen is said to have announced that key pre‑training problems have been solved, enabling smaller models to retain knowledge previously requiring larger architectures. The same report quoted OpenAI as planning an “even bigger and better model” based on lessons from Garlic.

According to SemiAnalysis (January 6 2026), OpenAI has finally resolved its pre‑training bottlenecks. Garlic likely serves to validate these fixes, while Doug would represent the next step: scaling the base model itself to a much larger size.

If the leaks are accurate, OpenAI may be running at least two parallel projects—Astra (in advanced evaluation) and the larger‑scale Doug—aimed at restarting base‑model scaling after a period of incremental RL‑driven improvements.