I'm LongbridgeAI, I can summarize articles.OpenAI is accelerating the development of a new model codenamed Garlic to address the competitive pressure brought by Google's Gemini 3. Internal testing shows that it has surpassed Gemini 3 and Anthropic's Opus 4.5 in programming and reasoning tasks. This model integrates Shallotpeat's pre-training repair solution, enabling more efficient knowledge injection at a smaller scale. The related technological iterations suggest that GPT-5.2 or GPT-5.5 may debut in early 2026. However, Garlic still needs to undergo further training and safety assessments, and its actual application performance remains to be observed
OpenAI is accelerating the development of a new generation large language model named Garlic to address recent competitive pressures perceived as lagging behind Google in the AI field, aiming to regain its technological leadership.
On December 2, sources revealed that OpenAI's Chief Research Officer Mark Chen stated in an internal communication last week that Garlic performed well in the company's internal evaluations, surpassing Google Gemini 3 and Anthropic's Opus 4.5 in key tasks such as programming and reasoning. The company plans to launch its related versions as soon as possible, which means GPT-5.2 or GPT-5.5 may debut in early 2026.
This acceleration comes as Google gains market attention with Gemini 3. Earlier reports indicated that OpenAI CEO Sam Altman had initiated a "red alert," calling for a comprehensive enhancement of ChatGPT's capabilities. As a first step, OpenAI is about to release a reasoning model that is "significantly ahead" of Gemini 3 in internal evaluations.
Technical Breakthroughs of Garlic Compared to Previous Models
Garlic is an independent model separate from Shallotpeat. Shallotpeat is a new large language model that OpenAI is developing, and Altman told employees in October that Shallotpeat would help OpenAI challenge Gemini 3. Garlic incorporates the error correction programs used during the pre-training process of developing Shallotpeat.
This is crucial because Google stated last month that it achieved new breakthroughs in pre-training while developing Gemini 3, a point that OpenAI's leadership has also acknowledged. Chen stated, during the development of Garlic, OpenAI addressed key issues encountered in pre-training, including optimizing its previously claimed "best" and larger pre-training model GPT-4.5, which has rarely been mentioned since its release in February of this year.
Essentially, Chen noted that these improvements mean OpenAI can inject equivalent knowledge into smaller models, which previously could only be achieved by developing larger models. Developing larger models is clearly more costly and time-consuming. Chen indicated that based on the experiences gained from developing Garlic, OpenAI has initiated the next phase of research and development for larger and more powerful models.
Multiple Tests Required Before Release
Garlic still needs to undergo multiple steps before its official release, including post-training (showing the model more carefully selected data to understand specific fields such as medicine or law, or how to better respond to chatbot users), other tests, and safety evaluations.
Currently, it remains unclear how developers will evaluate Garlic in practical applications. As has been repeatedly observed, performing excellently in internal evaluations is one thing, while maintaining stable performance in real-world scenarios such as programming and data analysis is another. The market response to GPT-5 serves as an example In the currently relatively adverse competitive environment, Chen and the OpenAI leadership also take on the role of internal morale builders, needing to convey confidence and boost morale within the team