OpenAI has unveiled its o3 model, which scored 88% on the ARC-AGI benchmark that measures AI capabilities. It outperforms previous versions such as o1, which achieved 32%.

The o3 model employs a technique called “test-time scaling,” which improves computational power during its inference phase.

While this method improves results, it increases operational costs.

François Chollet, who developed the ARC-AGI benchmark, said that the highest-scoring version of o3 required resources amounting to over US$10,000 per task.

Recent OpenAI developments