Highlights:
- ChatGPT Pro offers unlimited access to OpenAI’s o1, its flagship LLM designed for reasoning tasks.
- ChatGPT Pro customers with more advanced needs can access a more powerful version of o1 called o1 pro mode.
Recently, OpenAI launched ChatGPT Pro, a new paid version of its chatbot that offers access to large language models designed for enhanced reasoning tasks.
The subscription costs USD 200 per month, ten times higher than the consumer-oriented ChatGPT Plus plan. This launch marks the beginning of a 12-day product announcement series that OpenAI previewed on recently. The company is expected to unveil its Sora video generation model during one of the upcoming announcements.
ChatGPT Pro offers unlimited access to OpenAI’s o1, the company’s flagship LLM for reasoning tasks. The subscription also includes o1-mini, a version that sacrifices some response quality for faster processing, as well as the more general-purpose GPT-4o.
OpenAI initially released o1 in preview last September. As the ChatGPT Pro plan rolls out recently, the company has made the LLM generally available, incorporating several enhancements.
The latest version of o1 delivers faster and more accurate responses than the preview release. According to sources, the model is 34% less likely to produce outputs with “major errors” when handling “difficult real-world questions.” Additionally, it now includes the ability to analyze images uploaded by users.
ChatGPT Pro customers with more advanced needs can access a more powerful version of o1 called o1 pro mode. According to OpenAI, this mode utilizes additional hardware resources to process prompts, leading to longer response times but a notable improvement in output quality.
OpenAI compared o1 pro mode and o1 using AIME 2024, a set of math competition questions. The pro version correctly answered 86% of the questions, outperforming the original by 9%. In another benchmark involving doctorate-level science questions, o1 pro mode achieved a 3% improvement over o1’s score.
OpenAI discovered that o1 pro mode performed only 1% better on a test of difficult programming questions. However, for simpler coding tasks, it generated 75% fewer errors than o1. OpenAI notes that these simpler tasks are more representative of the programming-related prompts the model is typically expected to handle.
Since o1 pro mode takes longer to generate responses, OpenAI has added a progress bar to ChatGPT Pro to track wait times. The chatbot also notifies users when the model completes a task.
OpenAI plans to gradually expand o1’s capabilities. It is developing a function calling feature that will enable the model to interact with external applications. Additionally, OpenAI plans to enable o1 to browse the web and process different file uploads, while ChatGPT Pro will receive “compute-intensive productivity features.”