TMTPOST -- OpenAI is reportedly developing a new large language model (LLM) codenamed Garlic to counter Google's recent AI advances, while CEO Sam Altman has declared a "code red" to marshal resources for improving ChatGPT, delaying initiatives including advertising plans, according to reports from The Information.
Related articles
AI Generated Image
Last week, OpenAI's chief research officer Mark Chen informed colleagues that Garlic was performing well in the company's evaluations compared to Google's Gemini 3 and Anthropic's Opus 4.5 in coding and reasoning tasks, The Information reported on Tuesday. Chen said OpenAI is looking to release a version of Garlic as soon as possible, potentially as GPT-5.2 or GPT-5.5 by early next year.
Separately, Altman told employees in an internal memo that OpenAI would delay other initiatives such as advertising, AI agents for shopping and health tasks, and Pulse personalized reports to focus on ChatGPT improvements, according to The Information on Monday. While OpenAI hasn't publicly acknowledged working on ads, it is testing different types including online shopping-related advertisements, the report said citing a person with knowledge of the plans.
The moves represent a reversal from three years ago when Google launched its own "code red" to respond to ChatGPT's threat to Google Search. While ChatGPT still leads with 800 million weekly users, Google said in October that Gemini reached 650 million monthly active users, up from 450 million in July.
New Model Incorporates Key Pretraining Advances
Garlic differs from Shallotpeat, another model under development that Altman told staff in October would challenge Gemini 3. The new model incorporates bug fixes OpenAI used developing Shallotpeat during the pretraining process, when an LLM learns from web data and other sources.
Chen said OpenAI solved key pretraining problems, improving upon GPT-4.5, which launched in February. The improvements allow OpenAI to infuse a smaller model with knowledge previously requiring a much larger model, presumably reducing costs and development time. OpenAI has moved on to developing an even bigger model based on lessons learned from Garlic, Chen said.
Garlic still requires post-training with curated data for specific fields, plus testing and safety evaluations before release. The model's real-world performance remains uncertain, as internal evaluations don't always translate to practical applications like coding or data analysis.
ChatGPT Performance Critical to Funding Plans
Altman specified that code red priorities include personalizing ChatGPT for its users, improving Imagegen image generation, enhancing model behavior to outperform competitors in public rankings like LMArena, boosting speed and reliability, and minimizing overrefusals when the chatbot incorrectly refuses benign questions.
ChatGPT's performance impacts OpenAI's ability to raise another $100 billion to weather significant cash burn. The company projected this summer that ChatGPT would generate $10 billion in revenue this year from subscriptions. $20 billion next year and roughly $35 billion in 2027, while burning tens of billions developing new technologies.
Altman said Monday that OpenAI plans to ship a new reasoning model next week that is "ahead of" Gemini 3 in internal evaluations, but the company needed more work on the ChatGPT "experience." ChatGPT leader Nick Turley said Monday the service handles 70% of the world's AI "assistant activity" and 10% of "search activity."
更多精彩内容，关注钛媒体微信号 （ID：taimeiti），或者下载钛媒体 App
