In December of the previous year, OpenAI launched its o3 "reasoning" AI model and collaborated with the creators of the ARC-AGI benchmark to demonstrate the capabilities of o3. However, several months later, the test results have been revised, and the current performance appears less impressive compared to the initial outcomes.

Last week, the Arc Prize Foundation, responsible for maintaining and managing ARC-AGI, updated its approximate cost estimation for o3. Initially, the foundation estimated that the best configuration of o3 tested (o3 high) would cost around $3,000 to solve a single ARC-AGI problem. Now, the Arc Prize Foundation believes the cost is much higher, potentially reaching $30,000 per task.

This revision is significant because it reveals that state-of-the-art AI models can be extremely costly for certain tasks, at least in their early stages. OpenAI has not yet priced o3 or released the model. However, the Arc Prize Foundation considers the pricing of OpenAI’s o1-pro model to be a reasonable reference.

Notably, o1-pro is OpenAI's most expensive model to date.

One of the co-founders of the Arc Prize Foundation stated that due to the amount of computation used during testing, they believe the real costs of o1-pro and o3 are closer. However, this remains an approximation, and before official pricing is announced, they have marked o3 as "preview" status on the leaderboard to reflect this uncertainty.

Considering the reported amount of computational resources used by the o3 high model, the high price is not surprising. According to the Arc Prize Foundation, o3 high uses 172 times more computational resources than o3 low (the lowest computational configuration of o3) when solving ARC-AGI problems.

Additionally, rumors about OpenAI considering launching expensive plans for enterprise customers have been circulating for some time. In early March this year, reports suggested that the company might charge up to $20,000 per month for specialized AI "agents," such as software development agents.

Some argue that even OpenAI's most expensive models cost far less than typical human contractors or employees. However, AI researcher Toby Ord noted in a post on the X platform that these models may not be as efficient. For example, o3 high needs to attempt 1,024 trials per task in ARC-AGI to achieve its optimal score.