OpenAI’s o3 “reasoning” AI model, showcased in December with ARC-AGI benchmark testing, faces revised cost estimates. Initially, the Arc Prize Foundation approximated a $3,000 cost per ARC-AGI problem for the o3 high configuration. These costs have been updated to potentially $30,000 per task.
This revision highlights the potentially high expense of sophisticated AI models for certain tasks, particularly in their early stages. OpenAI has not yet released or priced o3, leading the Arc Prize Foundation to use o1-pro model pricing as a comparison.
Mike Knoop, co-founder of the Arc Prize Foundation, stated that “o1-pro is a closer comparison of true o3 cost … due to amount of test-time compute used,” emphasizing it as a proxy until official pricing is announced. The o3 remains labeled as “preview” on their leaderboard, reflecting this pricing uncertainty.
The potential high cost of o3 high aligns with its significant computing resource consumption. According to the Arc Prize Foundation, o3 high utilized 172 times more computing power than o3 low, the lowest-computing configuration of o3, when tackling ARC-AGI.
Rumors suggest OpenAI is considering high-priced plans for enterprise customers. In early March, The Information reported potential charges of up to $20,000 per month for specialized AI “agents,” such as a software developer agent.
While some argue that OpenAI’s models may still be cheaper than human contractors, AI researcher Toby Ord noted on X that models may not be as efficient. For instance, o3 high required 1,024 attempts per task in ARC-AGI to achieve its best score.
