OpenAI Launches New AI Model “o1” with Advanced Reasoning Capabilities

Created on 12 September, 2024 • News • 808 views • 5 minutes read

OpenAI Unveils o1 AI Model, Enhancing Problem-Solving and Reasoning Abilities

On September 12, 2024, OpenAI introduced its highly anticipated new AI model, the o1, as part of its ongoing series of AI systems designed to enhance reasoning capabilities. Internally codenamed “Strawberry” during development, the o1 model is now being released alongside a smaller, more affordable variant called o1-mini. These models are the first in a planned series of AI systems focused on tackling complex tasks that require logical reasoning, coding, and multistep problem-solving—tasks previously beyond the capabilities of its predecessors like GPT-4o.

OpenAI’s release marks a significant leap in AI technology, offering a preview version of o1 for ChatGPT Plus and Team subscribers starting today, while enterprise and educational users are slated to gain access early next week. Developer access to the o1 model through OpenAI’s API, however, comes with a steep price, with costs significantly higher than those for the widely used GPT-4o.

Breaking Ground in Reasoning

The o1 model stands out from previous AI iterations, including GPT-4o, due to its refined reasoning capabilities. Unlike its predecessors, which were trained to mimic patterns from vast datasets, o1 has been specifically engineered to solve problems on its own using reinforcement learning. This new training method allows o1 to process queries by breaking them down step by step, much like human reasoning.

According to OpenAI’s research lead Jerry Tworek, o1 has been trained with a novel optimization algorithm and a specialized dataset designed to challenge its reasoning abilities. This new approach makes o1 more adept at solving complex coding problems, advanced mathematics, and scientific tasks. Despite its advancements, Tworek acknowledged that the model has its limitations, particularly when it comes to "hallucinations," or generating incorrect or misleading information.

“We have noticed that this model hallucinates less,” Tworek said. “But the problem still persists. We can’t say we solved hallucinations.”

The improvements in reasoning are evident in practical tests. OpenAI’s chief research officer, Bob McGrew, highlighted how o1 outperformed GPT-4o in mathematics and logic-based challenges. For instance, o1 scored 83% on a qualifying exam for the International Mathematics Olympiad, compared to GPT-4o’s 13% success rate.

“We’ve been spending many months working on reasoning because we think this is actually the critical breakthrough,” McGrew said. “Fundamentally, this is a new modality for models in order to be able to solve the really hard problems that it takes to progress toward human-like levels of intelligence.”

Advanced Coding and Multistep Problem-Solving

One of the key areas where o1 shines is its ability to handle complex coding tasks. OpenAI claims the model performs exceptionally well in coding competitions such as Codeforces, where it reached the 89th percentile of participants. The model is also better at explaining its reasoning as it works through problems, offering users a clearer understanding of how it arrives at solutions.

OpenAI designed the o1 model to process tasks similarly to human thought processes. In demos, the model demonstrated the ability to break down puzzles and mathematical problems step by step, often using phrases like “I’m thinking through” or “Let me see” to simulate human-like reasoning. However, OpenAI is quick to note that while the model may feel more human in its reasoning approach, it is not actually “thinking” like a human.

During a demonstration, the o1 model was asked to solve a complex puzzle involving the ages of a prince and princess. The model took around 30 seconds to deliver the correct answer while showing each step of its reasoning, mimicking a process of logical deduction.

This new reasoning ability makes o1 well-suited for industries that require complex problem-solving, such as healthcare, engineering, and advanced scientific research. OpenAI has suggested that o1 could help annotate complex cell sequencing data in biology or assist physicists in generating intricate mathematical formulas for quantum optics.

Pricing and Accessibility

While the o1 model brings significant advancements in reasoning, it comes with a high price tag. In the OpenAI API, the o1-preview model costs $15 per million input tokens and $60 per million output tokens—far more expensive than GPT-4o, which is priced at $5 and $15, respectively, for the same token counts.

Despite the cost, OpenAI is making the o1-mini model more accessible. This smaller version of the model is designed for faster, more affordable responses, making it ideal for users with simpler needs or those focused on STEM-related queries. OpenAI has stated that o1-mini will be made available to all free-tier ChatGPT users, though no specific release date has been set.

Starting today, ChatGPT Plus and Team subscribers will be able to try o1-preview and o1-mini with a limit of 30 and 50 messages, respectively. Enterprise and educational users will gain access next week, while developers will be able to prototype with both models through OpenAI’s API, though with certain rate limits.

Limitations and Future Developments

While o1 excels in reasoning and problem-solving, it lacks some of the functionalities present in GPT-4o. For example, the o1 model does not support browsing the web, processing files or images, or handling custom system messages. These omissions mean that o1 is not a direct replacement for GPT-4o in many applications but is instead positioned as a complementary tool for solving more complex, reasoning-intensive tasks.

Despite its limitations, OpenAI believes the o1 model represents a significant step forward in AI development. As the company continues to refine the model and gather feedback, future updates to o1 are expected to introduce new features and improvements. OpenAI is also planning to enhance the o1 series with capabilities such as browsing and file uploading, making the models more versatile and practical for a wider range of applications.

“We’re resetting the counter back to 1 with this model,” McGrew said, referring to the model’s naming convention. “It’s the first step toward a new class of AI capabilities, and there’s more to come.”

The release of o1 and its mini variant marks a pivotal moment in the AI landscape. With its enhanced reasoning abilities, the o1 model is poised to make significant contributions to fields that require complex problem-solving, from coding to advanced scientific research. While it is not yet a complete replacement for GPT-4o in many use cases, o1 represents the beginning of a new era in AI development.

As OpenAI continues to push the boundaries of what AI can achieve, future updates to the o1 series promise to bring even more powerful capabilities, further closing the gap between human and machine intelligence.