OpenAI’s o3 Model Estimated to Cost $30,000 per Task, Far Surpassing Initial Projections

OpenAI’s o3 “reasoning” AI model, which was first unveiled in December 2025, may be far more costly to operate than initially anticipated. The model, which was developed in collaboration with the creators of ARC-AGI, a benchmark designed to test highly capable AI, has seen its computing cost estimates revised significantly.

According to a recent update from the Arc Prize Foundation, which administers the ARC-AGI benchmark, the cost of running o3 high, the best-performing configuration of the model, may be as high as $30,000 per task. The new pricing stands at $14,000 per task compared to the initial $3,000 per task estimate which creates doubts about using such models economically for specific tasks.

The revised estimates demonstrate how expenses for leading AI models in development have grown significantly. OpenAI has not revealed o3 pricing or released the model officially yet but the Arc Prize Foundation uses OpenAI’s most expensive AI o1-pro model pricing to estimate potential o3 costs.

Mike Knoop, co-founder of the Arc Prize Foundation, stated, “We believe o1-pro is a closer comparison of true o3 cost due to amount of test-time compute used. But this is still a proxy, and we’ve kept o3 labeled as preview on our leaderboard to reflect the uncertainty until official pricing is announced.”

The high expense of o3 stems from the enormous calculating power necessary to perform its computational tasks. The high consumption of computing resources by the o3 high variant exceeded the low configuration of o3 by 172 times, according to the Arc Prize Foundation report.

Enterprise customers of OpenAI face uncertainty about the detailed pricing structure that has been speculated to reach monthly fees of up to $20,000. Early assessments indicate that OpenAI plans to bill its enterprise customers $20,000 each month for their specialist AI “agents,” which can include software developer agents and others.

OpenAI’s expert Toby Ord expressed doubts about the model’s effectiveness despite claims that the pricey models remain less costly than traditional human service contracts. The specific model o3 high needed 1,024 tries for each individual task on the ARC-AGI platform to perform at its peak, which could create questions about its practical expense for businesses.

As OpenAI prepares to unveil the official pricing for o3, the revised estimates suggest that the company may face significant challenges in balancing the performance of its AI models with the cost of running them at scale.