The O3 OpenAi model may be more expensive than estimate in the first place

When Openai revealed it O3 “Thinking” Artificial Intelligence Model In December, the company has made a partnership with ARC-AGI creators, a standard designed for the very capable artificial intelligence test, to display the capabilities of O3. After months, the results were reviewed, and now it looks less impressive than it was at first.

Last week, the Arc Prize Foundation, which retains and runs the ARC-AGI, has updated O3 approximate computing costs. The Foundation has originally estimated that the best performance for the tested O3, O3 High, costs about $ 3,000 to solve the single ARC-Eagi problem. Now, the Arc Award Foundation believes that the cost is much higher – Perhaps about $ 30,000 per mission.

The review is noticeable because it shows the cost of the most expensive artificial intelligence models today at certain tasks, at least early. Openai was not yet pricing O3 – or even divorced. But the Arc Award Foundation believes Openai’s O1-PRO model Pricing is a reasonable agent.

For context, the O1-PRO is the most expensive OpenAI to date.

“We believe that O1-PRO is a more closer comparison of the real O3 cost (…) because of the extent of the test time used,” said Mike Konop, one of the co-founders of the ARC Award, for Techcrunch. “But this is still an agent, and we have maintained O3 in the name of an inspection on our leader’s board to reflect uncertainty until the official pricing is announced.”

The high price of O3 High will not be not possible, given the amount of computing resources used by the model. According to the ARC Arc, O3 HD uses 172X more than O3 LOW, which is the lowest composition for the O3 Pregnancy, to treat ARC-AAGI.

Moreover, the rumors were flying for some time about the price plans that Openai was thinking about offering to institutional customers. In early March, information I mentioned The company may plan to charge up to $ 20,000 per month for specialized “agents” of specialized artificial intelligence, such as software developer.

Some may argue that even the most batch Openai models will cost well in light of what a typical contractor or human employee will require. But as an artificial intelligence researcher Toby Ord He indicated in a post on XThe models may not be effective. For example, high O3 needed 1024 attempts at each task in the ARC-Eagi to achieve its best degree.