On Friday, OpenAI revealed benchmark scores for its latest series of artificial intelligence (AI) models known as the o3 series. This new line is positioned as a successor to the recently launched o1 series, with OpenAI asserting that the o3 models greatly enhance performance compared to their predecessors. The o3 series consists of two models: the o3 and the o3-mini, both designed for advanced reasoning tasks. Currently, these models are undergoing public safety testing, with plans to release the o3-mini for public use in early 2025. The announcement follows closely on the heels of Google’s unveiling of its Gemini 2.0 Flash Thinking Mode AI model.
OpenAI Unveils Benchmark Scores for the o3 Series AI Models
During the final day of a 12-day shipping announcement, OpenAI shared news of the o3 series, named after the skipped o2 designation, which was dropped due to trademark conflicts with the UK telecommunications company Telefonica, as noted by CEO Sam Altman in a live stream.
OpenAI clarified that the o3 models are not yet available for public use. Instead, the company has initiated early access for select external researchers to facilitate public safety testing. Those interested in contributing to this process can apply here, with the application window closing on January 10.
The company emphasized that the o3 series AI models offer significantly enhanced capabilities compared to prior iterations, particularly in tasks related to coding, mathematics, and natural language processing, although an exhaustive range of functionalities remains undisclosed.
Additionally, OpenAI disclosed preliminary benchmark results from internal testing of the o3 models, reporting that the o3 achieved scores of 71.7 percent on the SWE-bench benchmark and 96.7 percent on the AIME 2024 benchmark, surpassing results from the o1 model. A complete benchmark assessment is anticipated once the models are made publicly accessible. The o3-mini is slated for launch in January 2025.
OpenAI Extends Unlimited Access to Sora for Paid Subscribers
In another development, Altman also announced a “day-13 bonus” for subscribers of ChatGPT Plus and Teams through an X (formerly Twitter) post, offering them unlimited access to Sora for the last two weeks of December. This move is made possible due to reduced server demand as many offices close for the holiday season.
Rohan Sahai, a product lead at OpenAI, further elaborated in a separate post, noting that Sora’s blend feature has been upgraded and users can now share AI-generated videos via shared links, even with those who do not have an account.
