OpenAI has officially unveiled the production version of its groundbreaking reasoning model, previously known as Project Strawberry, now titled “o1.” Alongside it is a “mini” variant, similar to GPT-4o, which promises quicker, more efficient interactions at the cost of a more limited knowledge base.
The o1 model showcases a variety of technical improvements. As the first model in OpenAI’s reasoning suite, o1 leverages human-like deduction to tackle complex inquiries spanning fields like science, coding, and mathematics, outperforming human speeds.
During trials, o1 was tasked with a qualifying exam for the International Mathematics Olympiad. In a striking comparison, while its predecessor, GPT-4o, solved only 13% of the posed questions accurately, o1 achieved an impressive 83%. Furthermore, in an online Codeforces competition, o1 secured a spot in the 89th percentile. It has also demonstrated the ability to confidently handle questions that previously stumped earlier models, such as determining which number is larger, 9.11 or 9.9. However, OpenAI emphasizes that this is merely a glimpse of the model’s full potential.
According to Jerry Tworek, the research lead at OpenAI, the new o1 “has been trained using a completely new optimization algorithm and a new training dataset specifically tailored for it.” By integrating reinforcement learning alongside a “chain of thought” approach, o1 is said to deliver more precise conclusions than prior iterations. Tworek noted that while the model has reduced instances of hallucination, “we can’t say we solved hallucinations.”
Starting today, both ChatGPT-Plus and Teams subscribers can experience o1 and its mini counterpart. Enterprise and Edu subscribers can expect access within the upcoming week. Additionally, o1-mini will eventually be made available to free-tier users, although a specific timeline has not been disclosed.
Developers should be aware that the API pricing for o1 represents a significant increase compared to GPT-4o. Accessing o1 will cost $15 for every million input tokens, in contrast to GPT-4o’s $5 rate, while output tokens will be priced at $60 per million tokens—fourfold higher than the previous model’s fee. This raises an intriguing question regarding o1’s ability to recognize the correct number of ‘R’s in the word “strawberry.”