The OpenAI o3 Model May Be More Expensive Than Expected

Published:April 3, 2025

Reading Time: 2 minutes

When OpenAI launched its o3 “reasoning” AI model in December, excitement was high. The company partnered with ARC-AGI, a benchmark designed to test advanced AI, to showcase o3’s abilities. But now, a reassessment of the costs suggests that running the model may be far pricier than first thought.

The Arc Prize Foundation, which oversees ARC-AGI, has updated its estimate. Originally, it predicted that running the best-performing version, o3 high, would cost around $3,000 per task. Now, that figure may be closer to $30,000.

This massive increase raises big questions. How affordable will advanced AI be? Can businesses and researchers justify the cost?

The Price of Smarter AI

The Arc Prize Foundation’s new calculations show a major gap in computing needs. o3 high uses 172 times more computing power than o3 low, the least resource-intensive version. That’s a huge difference. More power means higher costs, and it’s not clear if the extra intelligence justifies the price.

Mike Knoop, a co-founder of the Arc Prize Foundation, shared some insight. He believes OpenAI’s o1-pro model is the best comparison for estimating o3’s cost. OpenAI hasn’t revealed o3’s official pricing yet, but o1-pro is already the company’s most expensive model. That sets expectations high.

A Glimpse Into OpenAI’s Pricing Strategy

Pricing speculation has been buzzing for months. In March, reports suggested OpenAI might charge businesses up to $20,000 per month for specialized AI “agents.” These AI tools could handle tasks like software development, but at a steep price.

For many businesses, AI’s potential to save time and money is appealing. However, if costs keep climbing, smaller companies may struggle to afford these tools.

Is the Cost Worth It?

Some argue that even OpenAI’s most expensive models could still be cheaper than hiring human experts. But AI researcher Toby Ord pointed out an important flaw. According to him, o3 high needed 1,024 attempts per task to get the best results in ARC-AGI tests. That’s a lot of effort. If AI takes that many tries to succeed, is it really efficient?

Businesses that rely on AI will need to weigh the benefits against the costs. High-powered AI models offer impressive capabilities, but they come with serious financial trade-offs.

Lolade

Contributor & AI Expert