The first batch of OpenAI's powerful models has arrived, with unlimited access to ChatGPT Pro's models, including the "smartest" o1

Author: Li Dan

Source: Hard AI

OpenAI CEO Sam Altman previewed the "big bomb" with the first batch released: OpenAI has launched a high-end GPT subscription package called ChatGPT Pro. Its fee is the highest among all of OpenAI's current products, offering unlimited access to all models under OpenAI, including OpenAI's strongest reasoning model o1 and an upgraded version of the o1 series.

On Thursday, December 5th, Eastern Time, OpenAI confirmed earlier online rumors and officially launched the package named ChatGPT Pro, with a monthly subscription fee of $200. OpenAI stated that subscribers to this package can access OpenAI's best models and tools on a large scale, including unlimited access to OpenAI's smartest model OpenAI o1, as well as the smaller models in the same series, o1-mini, GPT-4o, and the advanced voice mode of ChatGPT, Advanced Voice.

In addition, the ChatGPT Pro package can also include a new version of o1 called o1 pro mode, which is unique to ChatGPT Pro. OpenAI claims that this new version uses more computing power, can think more deeply, and provides better answers to the most difficult questions. It hopes to add more powerful compute-intensive productivity features to ChatGPT Pro in the future.

Altman introduced on social media that OpenAI has two new moves this Thursday: one is to include o1 in the ChatGPT Plus package for a monthly fee of $20, and the other is to launch ChatGPT Pro for a monthly fee of $200, which allows subscribers unlimited use of the models, including the even smarter o1.

o1 pro mode is stronger and more reliable than o1 in mathematics, science, and coding

OpenAI believes that ChatGPT Pro provides researchers, engineers, and others who use research-level intelligence daily with a new way to enhance their productivity, allowing them to be at the forefront of artificial intelligence (AI) advancements.

OpenAI specifically introduced o1 pro mode, stating that ChatGPT Pro offers "a version of OpenAI's smartest model," which "can think for longer periods, resulting in the most reliable responses." In evaluations by external expert testers, o1 pro mode can produce more reliable, accurate, and comprehensive responses, especially in fields such as data science, programming, and case law analysisThe chart below shows that in challenging machine learning (ML) benchmarks such as mathematics, science, and coding, the performance of o1 pro mode surpasses that of o1 and o1-preview. In mathematics, o1 pro mode scored 86, while o1 and o1-preview scored 78 and 50, respectively. In coding, o1 pro mode scored 90, with o1 and o1-preview scoring 89 and 62, respectively. In PhD-level science questions, o1 pro mode scored 79, while the latter two scored 76 and 74, respectively.

To highlight the main advantage of o1 pro mode—higher reliability—OpenAI has also raised the evaluation threshold, requiring that only when all four attempts correctly answer the questions can it be considered that the model has solved the problem, rather than just answering correctly once. Even under this high standard, the performance of o1 pro mode is significantly better than that of o1 and o1-preview.

The chart below shows that under the requirement that all four answers must be correct, the advantage of o1 pro mode over o1 and o1-preview is even greater. In mathematics, o1 pro mode scored 80, while o1 and o1-preview scored 67 and 37, respectively. In coding, o1 pro mode scored 75, with o1 and o1-preview scoring 64 and 26, respectively. In PhD-level science questions, o1 pro mode scored 74, while the latter two scored 67 and 58, respectively.