OpenAI readies 'Garlic' AI model to rival Google Gemini 3
What's the story
OpenAI is said to be working on a new large language model (LLM) codenamed "Garlic." The model is being developed as a competitor to Google's all-new Gemini 3 and Anthropic's Opus 4.5, especially in advanced reasoning and coding capabilities. According to The Information, early internal tests show Garlic performing strongly and it could debut as GPT-5.2 or GPT-5.5 by early 2026.
Project details
Garlic's development amid AI competition
The Garlic project comes as the AI race heats up with Google's success with Gemini 3. OpenAI's Chief Research Officer Mark Chen has reportedly told his team that Garlic has shown "strong performance" across several benchmarks, including reasoning and programming, where Google and Anthropic currently lead. To reclaim its position in the AI race, OpenAI CEO Sam Altman has declared a "code red" within the company to improve ChatGPT.
Launch plans
Garlic's launch timeline and development process
Despite no official comment from OpenAI, insiders say the company is fast-tracking Garlic's release with an early 2026 rollout in mind. The model builds on lessons learned from Shallotpeat, an earlier in-house model that Altman mentioned to employees in October. Unlike Shallotpeat which was aimed at challenging Gemini 3, Garlic incorporates bug fixes and refinements from that project especially during its pretraining phase.
Model advancements
Garlic's efficiency and performance
Chen claims Garlic is a major improvement in pretraining efficiency. He said the team has been able to "infuse a smaller model with the same amount of knowledge" that previously required a much larger one. This means Garlic could deliver GPT-4.5-level performance at lower cost and faster speed, giving OpenAI an edge over Google, which has been boasting similar improvements with Gemini 3's training process.
Model progress
Garlic's pretraining results and future plans
Garlic has already surpassed OpenAI's "previous best" pretraining results and overcome major technical bottlenecks that plagued GPT-4.5, which was launched earlier this year. The company is now confident it can create smaller yet more powerful models without increasing training costs. Before its launch, Garlic will undergo post-training with specialized datasets as well as safety testing and evaluation.