OpenAI is facing a sharper competitive moment as Google’s Gemini 3 gains ground across AI benchmarks. According to a report from The Information, Sam Altman told staff in an internal memo that Google’s recent progress could "create some temporary economic headwinds for our company" and that he expects "the vibes out there to be rough for a bit."
The memo points to a shift in the AI race: OpenAI’s lead over rivals such as Google and Anthropic appears to be narrowing. The company’s response includes a new language model codenamed "Shallotpeat," which is being developed as OpenAI tries to address problems in pre-training.
Why Gemini 3 changed the conversation
Google’s Gemini 3 is described in the source article as having put Google in the number one spot across nearly all benchmarks. That matters because benchmarks are one of the public signals the AI industry uses to compare frontier model performance, even when they do not capture every practical detail of how a model behaves in the real world.
The report says the memo followed claims that Google had developed a new AI that seemed to have leapfrogged OpenAI in its development methodology. The central point is not only that Google produced a strong model, but that its progress suggests a meaningful technical advantage in a part of model development OpenAI has reportedly found difficult.
Altman’s comments do not suggest panic. They do, however, acknowledge pressure. By warning about possible temporary business effects and a rougher external mood, he appears to be preparing staff for a period in which OpenAI may no longer be seen as clearly ahead in every major area.
Pre-training is back at the center
The key technical issue in the source is pre-training. This is the stage where an AI model learns from large amounts of data before later steps shape how it responds, reasons, and follows instructions. The article says this phase had seemed to be approaching its limits, but Google’s success shows useful advantages can still be found there.
In the internal note, Altman said Google has "been doing excellent work recently," especially in pre-training. That acknowledgement is important because it identifies the area where Google’s recent progress appears to have been strongest.
For OpenAI, this is a sensitive point. The source says the company has reportedly struggled to make progress in pre-training and has responded by placing more emphasis on "reasoning" models. It also says these issues showed up during GPT-5 development, where optimizations stopped working as the model was scaled up.
That creates a clear strategic tension. Reasoning models remain important, but the report suggests OpenAI cannot afford to let pre-training become a weak foundation. If Google has found gains in that earlier phase, OpenAI needs to improve there as well rather than relying only on later-stage model behavior.
What Shallotpeat is meant to fix
OpenAI’s answer, according to the report, is a new language model codenamed "Shallotpeat." Altman told staff the company would catch up, and a person familiar with the matter said the model is aimed specifically at fixing bugs that have occurred in the pre-training process.
The codename appears to reflect the difficulty of the task. The source notes that shallots do not grow well in peat because the soil is not ideal. In that sense, the name seems to point toward a model built to perform better on difficult training ground by addressing weaknesses in pre-training basics and data.
The practical stakes are straightforward:
- Google’s Gemini 3 has created new competitive pressure for OpenAI.
- Pre-training has become a visible area of comparison between the companies.
- OpenAI’s Shallotpeat effort is reportedly focused on bugs in that process.
- Altman is asking staff to stay focused despite short-term pressure.
The source does not provide a release date, performance target, or public product plan for Shallotpeat. What it does provide is a picture of a company trying to repair a technical bottleneck while staying committed to larger research goals.
Short-term pressure versus ambitious bets
Altman’s memo also frames the situation as a test of focus. He said OpenAI should continue to make "very ambitious bets," even if that means the company gets "temporarily behind in the current regime." The source says those bets likely include automating AI research itself to accelerate breakthroughs.
That framing matters because it separates two timelines. On one timeline, OpenAI has to respond to Google’s immediate momentum with Gemini 3 and show that it can improve its model development process. On the other, Altman is telling the research organization to keep aiming at superintelligence rather than reacting only to the latest competitive signal.
He wrote, "We need to stay focused through short-term competitive pressure," and added that it is "critically important" for most of the research team to remain focused on achieving superintelligence. Those lines make the internal message clear: Google’s progress is serious, but OpenAI does not want the response to become purely defensive.
The challenge is that both needs are connected. If OpenAI wants to keep making ambitious bets, it also has to maintain confidence in the core methods used to build stronger models. Shallotpeat, as described in the report, is part of that effort: a project aimed at improving the training foundation at a time when Google has shown that pre-training can still produce meaningful gains.
What to watch next
The source article leaves several major questions unanswered, including how far along Shallotpeat is and whether it will visibly change OpenAI’s position against Gemini 3. It also does not say how OpenAI will measure whether the pre-training bugs have been fixed.
Still, the direction is clear. Google’s recent AI progress has forced OpenAI to confront a weaker point in its development pipeline, while Altman is urging staff to stay committed to long-range research ambitions. The result is a more competitive AI race, with pre-training again treated as a crucial battleground rather than a solved stage of model building.