OpenAI’s o3 Model Might Be More Expensive Than Expected

When OpenAI introduced its o3 AI model in December, it was praised for its advanced reasoning capabilities. To highlight its potential, OpenAI collaborated with the team behind ARC-AGI, a benchmark designed to test highly capable AI models. At first, the results looked promising. But now, months later, a new analysis suggests that running o3 might be far more expensive than originally thought.

The Arc Prize Foundation, which oversees ARC-AGI, recently updated its cost estimates for o3. Initially, they believed that the most powerful version of o3, called "o3 high," would cost about $3,000 to solve a single ARC-AGI problem. However, after further evaluation, they now estimate the cost could be closer to $30,000 per task—a staggering increase.

This revised estimate raises concerns about how costly advanced AI models can be, especially in their early stages. OpenAI hasn’t officially announced the pricing for o3 yet, but some experts believe it may follow the pricing structure of o1-pro, OpenAI’s most expensive model so far.

Mike Knoop, co-founder of the Arc Prize Foundation, told TechCrunch that o1-pro might be a more accurate reference point for estimating o3’s cost due to the amount of computing power required. However, he also emphasized that the pricing remains uncertain until OpenAI provides official details.

It’s not surprising that o3 high is expensive to operate. According to the Arc Prize Foundation, this version of the model used 172 times more computing power than o3 low, the least resource-intensive version of o3. That level of processing demand naturally drives up costs.

Meanwhile, speculation is growing about OpenAI’s pricing plans for enterprise customers. Reports suggest the company may charge up to $20,000 per month for specialized AI agents, such as those designed for software development.

While some argue that even OpenAI’s most expensive models could still be cheaper than hiring a human professional, others question their efficiency. AI researcher Toby Ord pointed out on social media that o3 high required 1,024 attempts to complete each ARC-AGI task at its best performance. If AI models need that many tries to get things right, their cost-effectiveness could be in doubt.

As AI continues to advance, one thing is clear—cutting-edge technology comes with a hefty price tag. The real question is whether businesses and researchers will find the cost of running o3 worth the investment.