The cost of achieving high benchmark scores in artificial intelligence has decreased significantly in recent years.

According to Epoch, OpenAI's o3 model required approximately $0.30 per question to reach a 75% score on GPQA Diamond. Less than a year and a half later, GPT-5.6 Luna achieved the same result for about $0.0004 per question. This represents a drop in costs by a factor of 725.

Epoch’s analysis suggests that since 2023, matching benchmark performance has become roughly 47% cheaper each quarter. This translates to about a 13-fold annual reduction across five key benchmarks.

Unlike previous studies focusing on token prices, Epoch examines the actual dollar amount needed to reach fixed benchmark accuracy, replaying benchmark transcripts under stricter token limits. The cost reduction rate for AI surpasses that of electricity, batteries, compute, or DNA sequencing.

📰 @aipost