OpenAI’s Reported “Code Red” and the Debate Over AI Pretraining

Google has attracted enormous attention for Gemini 3 and its TPU infrastructure. In response, reports say Sam Altman declared a “code red” at OpenAI, while The Information reports a new model in development called Garlic. In December 2025, I see Google as unusually well positioned: it has frontier models, infrastructure, custom chips, substantial revenue, researchers and data.

SemiAnalysis argues that OpenAI’s leading researchers had not completed a successful full-scale pretraining run broadly deployed as a new frontier model since GPT-4o in May 2024. That is the publication’s characterization of a complex technical history, not independently established proof that OpenAI stopped improving models. It raises the question of whether simply scaling pretraining has become harder.

OpenAI co-founder Ilya Sutskever, now at Safe Superintelligence, discussed this question with Dwarkesh Patel. In the clip he describes 2012–2020 as an age of research and 2020–2025 as an age of scaling, and questions whether 100 times more scale would transform results. My interpretation is that we need new research ideas and algorithms, not only more of the same computation.

Andrej Karpathy has also raised questions about scaling, and Yann LeCun has long argued that current language models have limitations. The debate even produced a meme contrasting reactions to similar claims from Sutskever and LeCun. These researchers do not necessarily hold identical positions, and the meme is not evidence that either has settled the issue.

Then Google released Gemini 3, which I see as a strong counterexample to the claim that scaling is finished. Perhaps Google’s methods or architecture helped it get past a constraint others encountered; many people point to its TPUs, custom chips developed over several years. The precise reasons for a model’s performance cannot be established from this comparison alone.

According to The Wall Street Journal, Altman told OpenAI employees to focus urgently on ChatGPT’s everyday experience, including personalization, speed, reliability and the range of questions it can answer. Other reporting said some initiatives would be delayed. I find the emphasis significant: product quality is about more than the model’s raw benchmark performance.

My view is that existing models are already capable enough for the vast majority of ordinary use cases, although I still welcome improvements in intelligence. For most people, the scaffolding around a model and the day-to-day experience matter enormously. My “99%” estimate is a judgment, not a measured figure.

Google has a strong model and growing usage; I cite Gemini’s rise from 450 million to 650 million active users and suggest OpenAI is approaching a billion. Those figures describe the period of the video and should not be assumed to use identical measures. Google also has multiple revenue streams and cash it can draw on while competing with OpenAI and Anthropic.

In a podcast clip, OpenAI chief research officer Mark Chen says the company has spent the previous half-year strengthening pretraining, including building a strong team and focusing on its important components. He says those efforts have produced artifacts that give him confidence OpenAI can compete with Gemini 3. That is an executive’s assessment, not an independently verified comparison.

Chen also says OpenAI sees considerable room left in pretraining despite claims that scaling is dead, and that its efforts have produced much stronger models. I take this as a sign that OpenAI is renewing attention to pretraining after a period in which reinforcement learning received enormous focus. Whether the reported gains carry through to a released product remains to be seen.

The Information reports that OpenAI is developing a model code-named Garlic to counter Google’s progress. According to its source, Chen told colleagues that Garlic was performing well against Gemini 3 and Claude Opus 4.5 on company evaluations involving coding and reasoning. These are reported internal results, and the article does not establish independent test conditions or final released performance.

The Information also reports that Altman described a separate reasoning model being prepared for launch as ahead of Gemini 3 in OpenAI’s internal evaluations. I wonder whether a year-end release might be called GPT-5.2 or GPT-5.5, but that is speculation. The outlet says Garlic is distinct from another development model called Shallotpeat and incorporates fixes used during its pretraining.

Reports also say OpenAI is delaying work on projects including advertising and shopping-related initiatives to focus on ChatGPT. I welcome a focus on the experience. Some people I know have switched heavily to Gemini, but they and I are not typical users; many people care more about the overall experience than a small increase in a benchmark.

Google can use its distribution through Search to grow Gemini, but ChatGPT was many people’s first exposure to this kind of AI. I think that familiarity and a strong product experience make most users less likely to switch for every incremental improvement, even as power users move to the current leader. OpenAI faces stronger competition, and consumers can benefit from that competition.