OpenAI has launched o3-mini, a refined version of its o3 reasoning model. This version is optimized to excel in areas such as science, mathematics, and coding, while maintaining low latency. Additionally, it operates with the search in ChatGPT, capable of finding updated information and providing links to relevant web content.
The o3-mini model is now available both in ChatGPT and through the API and is accessible to users of the Plus, Team, and Pro plans. OpenAI has announced that it will become available in the Enterprise plan this coming February. Furthermore, all paying users will have access to o3-mini high, a version offering greater precision and intelligence, albeit with slightly slower response times.
In addition to this, users of the free ChatGPT plan will be able to try o3-mini by selecting the “Reason” option next to the chatbot’s text input. This is a notable development since it is the first time that OpenAI has activated a reasoning model in its free version.
The developer asserts that o3-mini exhibits performance similar to that of the o1, the predecessor reasoning model of o3, and o1-mini across several parameters. It is noteworthy that o3-mini has been developed to excel in STEM reasoning (science, technology, engineering, and mathematics), akin to the o1 model, but with enhanced effectiveness.
“With a moderate reasoning effort, o3-mini matches the performance of o1 in mathematics, coding, and sciences, while offering faster responses. Evaluations conducted by expert reviewers demonstrated that o3-mini produces more accurate and clear responses, with stronger reasoning capabilities, compared to OpenAI o1-mini,” explains OpenAI.
During these evaluations, it is highlighted that 56% of the time, reviewers preferred the responses provided by o3-mini over those generated by o1-mini. Additionally, “they observed a 39% reduction in significant errors in difficult real-world questions.”
OpenAI has announced that o3-mini will replace o1-mini in the ChatGPT model selector. This model will offer higher speed limits and lower latency. Furthermore, the limit for the Plus and Teams plans is increased from 50 messages/day with o1-mini to 150 messages/day with o3-mini.
“While OpenAI o1 remains our most comprehensive general knowledge reasoning model, OpenAI o3-mini provides a specialized alternative for technical domains requiring precision and speed,” they affirm.
We wanted to test the new o3-mini model by OpenAI by comparing it with the GPT-4 version, the most powerful non-reasoning specialized conversational generative AI from the company. For this purpose, we posed the following question to both models:
“Imagine you have three boxes: one contains only apples, another only oranges, and the third a mix of both, but all are incorrectly labeled. How can you identify the actual contents of each box by taking a single fruit from one of them?”
In the case of GPT-4, the AI provides a short answer presenting one of the possibilities as an example of problem-solving.
Response of ChatGPT4
In contrast, using the o3-mini model, the AI generates a much more elaborate response that not only presents the steps it follows during reasoning but also provides the different possibilities for solving the problem.
Response of ChatGPT o3
Photo: GPT4
Your email address will not be published. Required fields are marked *
Δ