OpenAI's o3 might cost $30,000 per task

Running OpenAI’s most powerful reasoning model might break the bank even faster than we thought.

Back in December, OpenAI stunned the AI world by flaunting its upcoming o3 model on ARC-AGI, a benchmark notorious for testing abstract reasoning. At the time, the organizers calculated that cracking a single task on the highest setting cost around $3,000.

Now, the bill has come due, and it is ten times higher. The Arc Prize Foundation just updated its estimates, revealing that running o3 high likely costs around $30,000 per single task.

Brute force meets diminishing returns

To be clear, OpenAI hasn't publicly released or priced o3 yet. So where does a $30,000 price tag come from?

The Arc Prize team used OpenAI's current top-tier tier, o1-pro, as a pricing proxy. Foundation co-founder Mike Knoop explained to TechCrunch that o1-pro provides a much more realistic picture because of the sheer volume of test-time compute o3 consumes.

The numbers back that up. Running o3 on high used a staggering 172 times more compute than its lower-tier o3 low sibling just to squeeze out those benchmark points.

And it turns out that "genius" reasoning involved a lot of desperate guessing. As AI researcher Toby Ord pointed out on X, o3 high needed up to 1,024 attempts per problem to produce its best score. That is not efficient problem solving; that is throwing computational horsepower at a wall until something sticks.

This also aligns with reports from early March by The Information, which revealed OpenAI has weighed charging enterprise clients up to $20,000 a month for specialized autonomous agents, such as coding assistants.

Why it matters

We think this estimate revision exposes the messy economic reality hiding behind recent reasoning benchmarks.

For businesses, a model that burns $30,000 on a single complex puzzle is completely off-limits for real-world production. Even if someone argues a top engineer costs more over time, an engineer does not need 1,024 failed drafts to solve one logic bottleneck. Unless you are a well-funded research lab or a massive tech giant, o3 at this tier is impossible to justify using.

For the broader industry, it proves that simply scaling "test-time compute" hits a painful economic ceiling very quickly. Brute-forcing 1,024 rollouts proves that current architectures still lack true intuition, relying instead on expensive Monte Carlo-style trial and error.

OpenAI pushed the boundaries of what models can do, but making an AI 172 times more expensive for incremental gains is a commercial dead end. The real race now isn't just about who builds the smartest reasoning agent, but who figures out how to deliver that logic without demanding a fortune for every solved prompt.