Inference Prices Plunge 214x in Two Years: The Collapse of Intelligence Cost

McDonaghMatthew · x · 2026-08-08

The author argues that we have been inside the technological singularity since 2022, highlighted by the dramatic collapse in the cost of intelligence. Historically bound to human labor, acquiring more analysis or engineering required hiring more people—a paradigm now breaking down.

According to data from Epoch AI, the lowest API price to achieve GPT-4-level performance (on the MMLU benchmark) plummeted from $37.50 per million weighted tokens in March 2023 to just $0.175 in February 2025.

This marks a 214x cost reduction for comparable intelligence in under two years. The article emphasizes this isn't just a single model getting cheaper, but the market learning to deliver capable intelligence at a fraction of the cost. Once inference becomes inexpensive enough to stop rationing, it will be embedded across every product, workflow, and decision.

Related event: Cost of GPT-4-Level AI Plummets 99.5% in Two Years(2 posts)→

Original post →

More from AGI Musings

AGI Musings channel →