The models, announced Friday, did so well on a prominent benchmark that some immediately questioned whether the company had in fact achieved the AI industry’s holy grail: artificial general intelligence, or AGI. It’s a much sought-after sci-fi milestone, when AI can perform at least as well as humans at certain tasks.

However, François Chollet, the benchmark’s creator, who recently left Google Research, responded with an unequivocal no. In a post on X after the OpenAI announcement, he said that while the new model is “very impressive and represents a big milestone on the way towards AGI,” there are still a “fair number” of simple tasks that o3 cannot solve. Still, Chollet added, “these capabilities are new territory, and they demand serious scientific attention.”

Meanwhile, the ARC Prize, a nonprofit initiative dedicated to advancing AGI research by challenging participants to solve the so-called ARC-AGI benchmark that OpenAI did so well on, also noted how o3 is a major advancement in AI. In a blog post, the organization said, “This is a surprising and important step-function increase in AI capabilities, showing novel task adaptation ability never seen before in the GPT-family models.”

The ARC-AGI benchmark rates an AI system’s ability to generalize and acquire new skills efficiently.

The announcement of o3 comes just a month after the AI community had grappled with speculation that the race to AGI—driven by Big Tech giants and well-funded startups like OpenAI and Anthropic—was facing significant hurdles. Concerns have centered on the so-called scaling laws, which suggest that AI performance improves predictably with increases in model size, dataset size, and computational power.

Critics argue that these predictions may no longer be holding up, and that improvements in AI are slowing. Inevitably, this raises questions about the limits of current AI approaches.

Still, the improvements represented by o3 show that those limits have not been reached yet. While it’s not AGI, the new models could lead to chatbots that handle more complex queries and solve problems step-by-step. AI search could evolve with improved semantic understanding (finding information not just based on keywords but on deeper meanings and relationships between concepts). The models may also generate more coherent stories and essays, as well as real-time problem solving.

The unveiling of the o3 model came on the 12th day of OpenAI’s product announcements, a marketing blitz timed to the holiday season. However, the “gift” can’t immediately be opened. The new models—o3 and o3 mini, a smaller, distilled model fine-tuned for particular tasks—are only available to safety researchers, who can apply to “explore and surface the potential safety and security implications of the next frontier models.”

OpenAI hasn’t confirmed a public release date.

OpenAI released its first model with what it called “enhanced reasoning capabilities,” called the o1 model, just three months ago. After the o3 announcement (which skipped o2 entirely), OpenAI researcher Noam Brown posted on X that “we have every reason to believe this trajectory will continue.”