OpenAI Secretly Building ‘Garlic’ to Take On Google Gemini 3 and Anthropic Opus 4.5
OpenAI is developing a new model, Garlic, aiming to surpass Google Gemini 3 and Anthropic Opus 4.5 in reasoning and coding.
The global AI race is intensifying, and OpenAI appears to be preparing a powerful new contender. According to a report from publication, the company is quietly building a next-generation model known as Garlic, designed specifically to challenge Google’s Gemini 3 and Anthropic’s Opus 4.5—both known for their advanced reasoning and coding capabilities.
Garlic is reportedly showing impressive early results. Internal testing suggests that the model is already performing strongly across key benchmarks, particularly areas where Google and Anthropic are said to have a lead. Sources familiar with the project say OpenAI may position Garlic as the successor to GPT-5.1, potentially launching it under the GPT-5.2 or GPT-5.5 branding as early as the first half of next year.
The report also notes that OpenAI CEO Sam Altman has raised the stakes internally, declaring a “code red” to accelerate improvements to ChatGPT and strengthen the company’s competitive edge. Altman is said to have told employees that OpenAI’s new reasoning model is already “ahead” of Google’s Gemini 3 in internal evaluations. While the company has not officially commented, insiders indicate that Garlic’s development and rollout are being fast-tracked to keep pace with rapid advances in the industry.
Garlic builds on the foundation laid by an earlier research project called Shallotpeat. Altman himself reportedly referenced this previous model in a meeting with employees in October. While Shallotpeat aimed to directly counter Gemini 3, Garlic takes the lessons and technical fixes from that effort and refines them further, especially in the crucial pretraining stage—the phase where a model learns to recognize complex patterns and relationships in vast datasets.
Mark Chen, OpenAI’s Chief Research Officer, has reportedly described Garlic as a major leap in training efficiency. According to Chen, the team succeeded in “infus[ing] a smaller model with the same amount of knowledge” that previously required a much larger architecture. This could allow OpenAI to deliver GPT-4.5-level performance at significantly lower cost and faster inference speeds. He added that Garlic has already exceeded the company’s “previous best” pretraining results and addressed long-standing technical bottlenecks encountered during the development of GPT-4.5.
This efficiency breakthrough mirrors claims made by Google about Gemini 3’s optimised training process—suggesting that the competition is no longer just about raw power, but about achieving more with less. Garlic’s progress may provide OpenAI with a stronger, more cost-effective foundation for future models, giving the company renewed momentum in the AI arms race.
Before Garlic is made public, it will undergo rigorous post-training refinement, safety evaluations, and specialized dataset tuning. Interestingly, sources say Garlic’s early success has already encouraged OpenAI to start work on its next successor model, which will build on the advances pioneered here.
If Garlic’s performance continues to meet expectations, OpenAI could regain significant ground in the global competition for AI dominance. With Google, Anthropic, and OpenAI all pushing toward more capable reasoning systems, the next generation of AI looks set to be more competitive—and more transformative—than ever.
