OpenAI’s o3 model might be costlier to run than originally estimated

Opena when it has unveiled O 3 “logic” AI model In December, the company and 3 were parted with the creators of a benchmark designed to test the extremely capable AI to display the capabilities of the agency and 3. A few months later, the results were fixed and they now look a little less impressive than they initially.

Last week, the Arc Prize Foundation, which maintained and operate the Arc-AGI, updated its estimated computing costs for 3. The organization originally assumed that it had spent about $ 3,000 to solve a single arc-AGI problem, and 3 high, 3 high, a single-arc-AGI problem. Now, the Arc Prize Foundation thinks that the cost is much higher – Perhaps about $ 30,000 per taskThe

The correction is significant because it is the most sophisticated AI models today depicted that at least for the first specific task. OpenAI may still pay the price of O3 – or even leave it. However the Arc Prize Foundation believes Openai O 1-Pro model A reasonable proxy to determine the price.

In the context, the most expensive model of the 1-Pro Openai is still the most expensive model.

“We believe in and a close comparison of 1-Pro true O3 expenditure […] Due to the amount of test-time computer, “Mike Nup, one of the co-founders of the Arc Prize Foundation, told Techcunch.” However, this is still a proxy, and we have labeled on our leadersboard to reflect u uncertainty until we and 3K official prices are determined. “

The amount of computing resources that the model uses is due to the amount of the resource and a high price for 3 high will not be out of question. According to the Arc Prize Foundation, O 3 1 Hi and 3 Low has used more than 172x computing, to deal with the Arc-AGI and the minimum-computer configuration of 3.

Furthermore, the Open Enterprise has been considering introducing customers for a long time. In early March, information Report So that the company can plan to charge up to $ 20,000 per month for a special AI “agent” like a software developer agent.

Some can argue that Openi’s prikeist models can even spend well under any ordinary human contractor or employee command. However as the AI ​​researcher Toby Ord Indicated in a post at XModels may not be more skilled. For example, O 3 Hi needs 1,024 attempts in each task of the Arc-AGI to achieve its best score.