The Arc Prize Foundation, which oversees the management of ARC-AGI, has significantly revised its cost estimates for OpenAI’s o3 “reasoning” artificial intelligence model. It appears the operating costs for o3 may be much higher than previously thought.
When OpenAI launched the o3 model last December, it collaborated with ARC-AGI developers to showcase the model’s impressive capabilities in solving complex problems. However, just a few months later, the estimated costs have undergone a dramatic change. The latest estimates from the Arc Prize Foundation suggest that the top-performing configuration of the o3 model, known as o3 high, could cost around $30,000 to solve a single ARC-AGI problem, a stark increase from earlier estimates of approximately $3,000.
This significant jump in cost highlights the high expenses that cutting-edge AI models may encounter for specific tasks. Although OpenAI has yet to publicly disclose pricing for the o3 model, the Arc Prize Foundation indicates that the pricing of OpenAI’s most expensive model, the o1-pro, could serve as a reference point.
Arc Prize Foundation co-founder Mike Knoop explained in an interview with TechCrunch, “We believe o1-pro is a closer comparison to the true cost of o3, as both models used similar computational resources during testing.” However, he emphasized that this is merely a reference value, and they label o3 as a preview model to reflect the uncertainty until official pricing is announced.
The high costs associated with o3 high are grounded in realities of resource usage. According to the Arc Prize Foundation, o3 high consumes 172 times the computational resources of the configuration with the least computational load, o3 low, when tackling ARC-AGI tasks. This extensive resource consumption contributes to the steep price tag.
Rumors have long circulated regarding OpenAI’s plans to charge premium fees for high-end, custom services aimed at enterprise clients. A report from The Information earlier this month suggested that the company might consider charging up to $20,000 per month for specialized AI “agent” services, such as software developer proxies.
While some argue that even the priciest AI models are still cheaper than hiring human contractors or employees, AI researcher Toby Ord pointed out in a post on X (formerly Twitter) that the efficiency of these models might not be as high as expected. For instance, o3 high required 1,024 attempts to achieve optimal results in the ARC-AGI tests.
As AI technology continues to evolve and commercialize, effectively controlling costs while maintaining high performance in models will become one of the industry’s prominent challenges.
