OpenAI introduces GPT-o1 with human-like reasoning and advanced capabilities

The model significantly outperforms older versions, scoring 83% in the International Mathematics Olympiad compared to GPT-4o’s 13%, and shows human-like hesitation patterns.
OpenAI has launched a safety system to stop its newest AI models from offering dangerous advice on chemical and biological threats, aiming to prevent misuse by bad actors.

OpenAI’s latest version of ChatGPT, GPT o1, a nomenclature indicative of resetting the counter clock to 1, and its less costly mini version, represents a watershed moment in the company’s LLM stockpile. Designed to replicate superhuman-level intelligence, the models can already answer questions a lot faster than humans. This series of models will be unlike previous ones. In responding to queries, they utilise a human-like ‘chain of thought’ processing combined with reinforcement learning on specialised datasets and optimisation algorithms. 

The model outperforms older models by a significant margin. For example, when tested against GPT-4o at the International Mathematics Olympiad, it scored 83 percent to GPT-4o’s 13 percent. What’s unique about the model is its ability to not only provide step-by-step reasoning for outputs but to show human-like patterns of hesitation during the process, ‘I’m curious about…’ and ‘Ok, let me see’ or ‘Oh, I’m running out of time, let me get to an answer quickly’. The new design has also resulted in a reduced occurrence of hallucinations. Yet, despite their many pros, the models have limitations. For instance, they cannot browse the internet, lack world knowledge, and cannot process files and images. 

According to the lead researcher on the project, Jerry Tworek, the next level is for the models to perform similarly to PhD students on challenging benchmark tasks in areas such as physics, chemistry and biology. He assures that the intention here is not to equate AI with human thinking but rather to illustrate the model’s ability to dive cognitively deep. For the company, reasoning is a step up from pattern recognition, which is the design model used with previous versions. Ultimately, OpenAI aims to develop a product that can make decisions and take action on behalf of humans, a venture estimated to cost a further $USD 150 billion. Removing the current kinks in the system will mean that the models can work on complex global problems we face today in areas such as engineering and medicine. 

More breakthroughs will also mean reduced access costs for developers and users. According to Chief Research Officer Bob McGrew, developer access to o1-preview is currently $15 per 1 million input tokens (chunks of text parsed by the model) and $60 per 1 million output tokens. GPT -o4 costs $5 per 1 million input tokens and $15 per 1 million output tokens.