AI gets about 13 times cheaper each year for the same result

Epoch AI's Luke Emberson and David Roodman put a number on something everyone in the field feels but rarely measures: how fast it is getting cheaper to reach a given level of AI performance. Their answer, across 2023 to 2026, is about 13 times cheaper per year, or roughly 47% every quarter. The clearest example in the piece: OpenAI's o3 scored 75% on GPQA Diamond for about $0.30 per question in January 2025. Eighteen months later, GPT-5.6 Luna matched that score for $0.0004 per question, a 725-fold drop.

The pace is not uniform. Math problems fell fastest at 50 to 52% per quarter, while game-style puzzles fell slowest at 39 to 43%. Costs also drop hardest right after a capability first appears, around 66% per quarter, then settle to about 32% once it is two years old. To give the trend context, Epoch compares it to other technologies: this decline runs about 4 times faster than DNA sequencing, 6 times faster than raw compute, and 54 times faster than electricity did.

Why it matters

If a capability is too expensive to ship today, the honest planning assumption is that it will be affordable within a year, so design for it now rather than ruling it out. The flip side: pricing a product on current inference costs is aiming at a target that keeps moving down, which reshapes margins and competition faster than most roadmaps assume.