ChatGPT's New Family: OpenAI o1 Unveils Advanced AI Reasoning

The Gist

OpenAI o1 launch: Advanced reasoning capabilities for complex problem-solving.
AI bargain? 80% cheaper for mini, with fast, cost-effective performance.
Reasoning feature: Reinforcement learning for more accurate, thoughtful responses.

OpenAI introduced OpenAI o1, the latest model family of the ChatGPT chatbot.

Marketers have long heard rumors about the model, inside the walls of OpenAI development known as “Strawberry." With the new release, OpenAI puts the rumors to rest with a truly valuable feature.

Open AI o1 launches with an advanced reasoning capability to perform more complex tasks, a remarkable advancement that has eluded its competitors in the development of AI.

What OpenAI Released to the World

OpenAI launched two versions of the o1 model, o1-preview and o1-mini. The o1-preview model is the standard large language model that powers ChatGPT. Its capabilities were meant to address complex steps for tasks such as programming and math.

The o1-mini is a smaller version designed specifically for the programming needs of compact applications and instances where broad world knowledge is not needed for the given use case. O1-mini is faster than o1-preview while being more cost-effective. OpenAI claims the mini is “80% cheaper than o1-preview” while still maintaining competitive performance in coding benchmarks.

Both models are available in ChatGPT Plus and Team. Developers can access o1 through OpenAI's API. Enterprise account users will have access to the new model family very soon.

The Reason Behind OpenAI o1’s Reasoning

A single word describes the main feature of OpenAI o1: reasoning. The training of the new model introduces reinforcement learning to perform complex reasoning. This means it “thinks” more carefully before it answers. The thinking means it can produce a long internal chain of “thought” before responding to the user prompt.

This technique is an advanced form of reinforcement learning, a training technique to teach a large language model how to think through the prompt details to craft a solution. LLMs have traditionally performed poorly on tasks involving common sense reasoning. In fact, many training models were designed with human intervention, called Reinforcement Learning with Human Feedback (RLHF), to fine-tune the model’s behaviors.

Today’s AI leaders have been researching ways to create more independent reinforcement learning in their models, but with the launch of o1, OpenAI is the first.

The reasoning in OpenAI o1 addresses how the model refines its response strategies to solve a problem, to better break down tricky steps into simpler ones, and to recognize its mistakes.

OpenAI refined its model training so that the model reasoning process mimicked a student’s learning ability to solve complex problems. When people usually think of a solution, they recognize mistakes being made and look at different approaches. The o1 model learns to try a different approach when the current one isn’t working. This process improves as the model further reasons. The longer o1 thinks, the better its reasoning on tasks becomes.

OpenAI tested the new model and found that it significantly outperforms its previous model GPT-4o on the majority of reasoning-heavy tasks. In fact, in confirmation tests on challenging tasks in physics, chemistry and biology, OpenAI discovered that o1 performs with reasoning comparable to that of PhD students.

OpenAI o1’s Features Have Limited Access

However, there are feature limitations in the new model compared to its predecessor. For starters, no files can be uploaded to the model. This means images and documents that have become a ChatGPT Plus staple are on hold — at least for now. Online searches are also eliminated, so users will have to pay special attention to the prompt response accuracy to know if the information is from the latest sources.

Finally, there are prompt limits in the number of messages that can be sent: up to 30 messages per week in the standard model, while the o1-mini has a 50-message per week capacity.

There are also some limitations on the API as well. Developers must be at a tier 5 to have access to the o1 API. OpenAI also encourages developers to review the API documentation to note what features cannot be currently called in the API portal.

The limits are understandable. Heavy account usage usually occurs after a significant AI model launch attracts public attention.

OpenAI noted that it plans to increase those rates “and enable ChatGPT to automatically choose the right model for a given prompt.”

OpenAI o1’s Advanced Reasoning Can Transform Marketing

The benefits of the enhanced reasoning capabilities in o1 will certainly help marketers whose products and services are engineered. These offerings and their associated operations are influenced by science, programming, mathematics and other similar fields.

The problem-solving bias of the model is well-suited for marketers from these fields and across several industries. The ability of o1 to think through problems and refine its reasoning could be beneficial in strategic planning, financial modeling or operational optimization. The result is optimized processes, solutions to logistical challenges or better data-driven marketing decisions.

OpenAI o1 represents just one view of the expected horizon for OpenAI. A flagship LLM code name Orion is still in development. There is still the high anticipation of SearchGPT, the search engine combined with ChatGPT. OpenAI closed the beta signup, but no news has appeared since the trial signup.

Learning Opportunities

Webinar