OpenAI working on a secret Garlic AI model to challenge Google Gemini 3 and Opus 4.5 in coding and reasoning

The race to build the most powerful artificial intelligence system has entered a new phase, with Microsoft-backed OpenAI quietly developing a large language model called Garlic. The model is being designed to rival Google’s Gemini 3 and Anthropic’s Opus 4.5, particularly in advanced reasoning and coding abilities, according to a report by The Information. Early internal tests suggest Garlic is performing strongly and could debut as GPT-5.2 or GPT-5.5 by early next year. The Garlic project comes amid growing competition following Google’s success with Gemini 3.

According to The Information, OpenAI’s Chief Research Officer, Mark Chen, told colleagues that Garlic had shown “strong performance” across multiple benchmarks, including reasoning and programming, where Google and Anthropic currently hold an edge. CEO Sam Altman has reportedly declared a “code red” inside the company to improve ChatGPT and reclaim OpenAI’s lead in the AI race. He told staff that OpenAI’s new reasoning model was already “ahead” of Gemini 3 in its own internal evaluations. Although OpenAI has not commented publicly, insiders say the company is fast-tracking Garlic’s release, aiming for an early 2026 rollout.

OpenAI's new model for smarter AI
Garlic reportedly builds on lessons from Shallotpeat, an earlier in-house model that Altman mentioned to employees in October. While Shallotpeat was designed to challenge Gemini 3, Garlic incorporates bug fixes and refinements from that project, particularly in the pretraining phase. This stage teaches a model to recognise relationships in massive datasets drawn from across the internet. According to Chen, Garlic represents a leap forward in pretraining efficiency. He told colleagues that the team had managed to “infuse a smaller model with the same amount of knowledge” that previously required a much larger one.

This advancement means Garlic could deliver GPT-4.5-level performance at lower cost and faster speed. The breakthrough comes as Google has been touting similar improvements with Gemini 3’s training process. OpenAI’s progress with Garlic could counterbalance that advantage and potentially give the company a more efficient path to future upgrades.

Chen said Garlic had already surpassed OpenAI’s “previous best” pretraining results and resolved key technical bottlenecks that affected GPT-4.5, which launched earlier this year. With these improvements, OpenAI is confident it can now develop smaller yet more capable models without inflating training costs. Before Garlic launches, it will undergo post-training with specialised datasets, along with safety testing and evaluation. Sources also claim the success of Garlic has already allowed OpenAI to begin working on an even more advanced successor model, building on the lessons learned during its development. If Garlic performs as strongly as early reports suggest, it could signal a major shift in the balance of power among AI giants. As OpenAI, Google and Anthropic push for dominance, the competition over the next generation of reasoning models is heating up, and Garlic may be OpenAI’s boldest move yet.