OpenAI has wrapped up its 12 days of pre-Christmas announcements by introducing two new AI models with enhanced reasoning capabilities: o3 and o3-mini. These models are direct successors to o1 and o1-mini, which were initially previewed in September and have since been fully integrated into ChatGPT.
Please follow us on Facebook and Twitter.
If you’re curious why OpenAI skipped from o1 to o3, the explanation is straightforward: it was to avoid trademark conflicts. It seems that Sam Altman’s company took precautions to steer clear of issues with Telefónica, which owns the O2 trademark.
Regarding the capabilities of o3 and o3-mini, OpenAI states that these new AI models function similarly to their predecessors. Instead of generating an immediate response to user queries, they can think and reason through their answers using various methods aimed at reducing hallucinations and improving the quality of results.
The o3 family of AIs employs a “private chain of thought,” allowing the models to take the necessary time to deliver thoughtful responses, OpenAI explained. Unlike o1, the new models offer three levels of thinking time: low, medium, and high. As expected, the low setting executes tasks more quickly but may be less accurate, while the high setting prioritizes accuracy at the cost of slower response times.
OpenAI takes an essential step toward AGI with the launch of o3 and o3-mini.
During the demo, OpenAI presented a few practical examples but mostly focused on showcasing the first evaluations of o3 and o3-mini. What stood out the most was that both models significantly outperformed o1 and o1-mini in the ARC-AGI benchmarks, which are specifically designed to measure AI capabilities against the human mind in the pursuit of artificial general intelligence (AGI).
In these tests, o3 scored 87.5%, far exceeding o1’s maximum score of 32%, and became the best-performing AI in these benchmarks. Does this mean OpenAI is close to achieving AGI? Not yet, but these models are making important progress toward that goal. Sam Altman’s team has committed to collaborating with ARC-AGI to develop new benchmarks and continue exploring the capabilities of their technology.
If you’re excited to try OpenAI’s new o3 and o3-mini models, you’ll need to wait a little longer. The company has stated that both are still in the security testing phase, so they won’t be available for public use for at least a few weeks. However, OpenAI has invited external experts and researchers to sign up for security testing on o3-mini, followed by o3.
Sam Altman mentioned that OpenAI aims to launch o3-mini by the end of January 2025, with o3 to follow later. We’ll be keeping an eye on the progress.