On Friday, Openai launched a new model for “logical”, O3-MINI, the latest in the family of thinking in the company.
Openai first inspected the model in December alongside a more capable system called O3, but the launch comes in a central moment of the company, whose ambitions – and challenges – are growing day by day.
Openai fights the perception that it waives the land of artificial intelligence to Chinese companies such as Deepseek, which claims Openai may have stolen its IP. However, the Chatgpt maker managed to beat dozens of developers, and he was trying to do so The beach of its relationship with Washington While following up an ambitious data center project at one time, it is also said to lay the foundations for one of the largest financing rounds by a technical company in history.
Which leads us to O3-MINI. Openai shows its new model as “strong” and “affordable”.
“The launch of the day represents an important step towards expanding the possibility of advanced artificial intelligence in the service of our mission,” Openai spokesman told Techcrunch.
The most efficient thinking
Unlike most large language models, thinking models such as O3-MINI are completely achieved by the facts of the right to provide results. This helps them avoid some of the pitfalls that usually rise. These thinking models take a little longer to reach solutions, but the comparison is that they tend to be more reliable-and if not perfect-in areas such as physics.
O3-MINI is set for STEM problems, specifically for programming, mathematics and science. Openai claims that the model is equal to the O1, O1 and O1-MINI family in terms of capabilities, but it works faster and costs less.
The company claimed that external laboratories prefer the O3-MINI answers over those in O1-MINI more than half the time. O3-MINI seems to have also committed 39 % of “major errors” on “difficult realistic questions” in A/b tests. Video of O1-MINI, and produced “more clear” responses with 24 % faster answers.
O3-MINI will be available to all users via ChatGPT starting from Friday, but users who pay the company’s Chatgpt Plus price and the team’s plans will get the maximum of 150 queries per day, while Chatgpt Pro subscribers will get unlimited access. Openai said that O3-MINI will come to Chatgpt Enterprise and Chatgpt Edu in a week (there is no word in Chatgpt government).
Users with distinctive Chatgpt plans can determine the O3-MINI using the drop-down menu. Free users can click or click the new “Reason” button in the chat bar, or get the “Reglerate” Chatgpt.
Starting on Friday, O3-MINI will also be available via API from Openai to identify developers, but at first you will not get support for image analysis. Devs can choose the “thinking voltage” level (low, medium or high) to get the O3-MINI “to think hard” based on their needs of their use and use.
The price of O3-MINI is $ 1.10 per million temporary stored input codes and $ 4.40 per million output symbols, as a million icons are equivalent to approximately 750,000 words. This is 63 % cheaper than O1-MINI, competitive with the prices of the thinking model in Deepsek R1. Deepseek receives $ 0.14 per million temporary stored input codes and $ 2.19 per million output symbols to reach R1 through its application programming interface.
In ChatGPT, O3-MINI has been assigned to an average thinking voltage, which Openai says, “A balanced comparison between speed and accuracy.” Users paid will have the option to choose “O3-MINI-He
Regardless of the version of O3-Mini Chatgpt users, the model will work with the Search to find updated answers with links to relevant web sources. Openai warns that the job is a “initial model” because it works to integrate research through thinking models.
“While O1 remains our broader general thinking model, O3-MINI provides a specialized alternative to technical domains that require accuracy and speed,” Openai wrote in a Blog Blog on Friday. “The O3-MINI version is another step on the Openai mission to push the limits of effective intelligence in terms of cost.”
Many warnings
O3-MINI is the most powerful Openai model yet, and it does not jump the R1’s Deepseek thinking in each standard.
O3-MINI beats R1 on AIME 2024, a test that measures the extent of understanding of models and response to complex instructions-but only with a high thinking voltage. It also outperforms R1 on the program that focuses on programming (by 0.1 Point), but again, only with high thinking voltage. In low thinking voltage, O3-MINI R1 falls behind GPQA Diamond, which tests models with physics, biology and chemistry at the PhD level.
To be fair, O3-MINI answers many inquiries at a low cost of competitiveness. In the post, Openai compares its performance with the O1 family:
“Through low-thinking voltage, O3-MINI achieves a similar performance with O1-MINI, while O3-MINI achieves a comparative performance with O1,” he writes Openai. “O3-MINI matches with an average thinking voltage O1 in mathematics, coding and science with faster responses. At the same time, with high thinking voltage, O3-MINI outperforms both O1-MINI and O1.”
It should be noted that the O3-MINI performance feature is minimal in some areas. In AIME 2024, O3-MINI O1 wins only 0.3 Celsius when he is set on high thinking voltage. And on GPQA Diamond, O3-MINI does not exceed the degree of high thinking.
Openai asserts that O3-MINI is “safe” or more safe from the O1 family, thanks to the red picking efforts and the “deliberative compatibility” methodology, which makes models “think” about Openai’s safety policy while responding to inquiries. According to the company, O3-MINI “greatly exceeds” one of the leading models in Openaii, GPT-4O, on “difficult safety and fracture assessments”.
