Prime Highlights:
- OpenAI’s o3 “reasoning” AI model will be higher in operation costs than anticipated.
- The Arc Prize Foundation raised o3 high cost per task from $3,000 to around $30,000.
Key Facts:
- o3 high used 172 times more computer resources than used by o3 low in ARC-AGI benchmark tests.
- Maximum scores for the model took 1,024 attempts per task.
Key Background
OpenAI has revealed in December its o3 “reasoning” AI model, showing how it performed through a collaboration with the ARC-AGI challenge developers. OpenAI originally estimated the price of completing a single ARC-AGI task using the best configuration, o3 high, as approximately $3,000. This recently blew up to perhaps $30,000 per task in new estimates.
This extreme transformation emphasizes the expense of running sophisticated AI models such as o3. The higher-costing costs are a result of the model’s heavy computational demands. In the course of ARC-AGI testing, o3 high used 172 times the amount of computing power than its lower-end counterpart, o3 low. In addition, in order to operate at best, o3 high needed 1,024 attempts per task, a testament to both its complexity and the related resource consumption.
OpenAI has not yet disclosed the official pricing for o3, but the Arc Prize Foundation suggests that OpenAI’s o1-pro model, the company’s most expensive model so far, could serve as a reference point. This suggests that deploying o3 may be a significant financial undertaking for potential users. Furthermore, there are hints that OpenAI will provide business customers with dedicated AI “agents” that will cost as much as $20,000 monthly. These are tendencies mirroring the growing cost model of high-end AI solutions.
Although models such as o3 are an incredible advance in the capability of AI, the corresponding costs of use can limit bulk adoption. Any companies considering models such as these must weigh heavily the potential bleeding-edge AI capability against the cost and computational complexity of implementing them.
