AI Performance Costs Falling ~47% Per Quarter, Faster Than Any Previous Technology

An Epoch AI report finds the cost of a fixed level of AI performance drops about 47% per quarter (~13-fold per year), with 75% on GPQA Diamond falling from $0.30 to $0.0004; MIT researchers independently estimate 5-10x annual declines.

An Epoch AI report by Emberson and Roodman, discussed on Marginal Revolution and Hacker News, finds that the cost of achieving a given level of AI performance fell about 47% per quarter over three years — roughly a 13-fold annual drop. As a concrete example, OpenAI's o3 cost about $0.30 per question to reach 75% on GPQA Diamond in January 2025, while GPT-5.6 Luna reached a similar score for about $0.0004 in mid-2026, a roughly 725-fold decline in under 18 months. The commentary argues frontier models are becoming both more capable and much cheaper to run at a fixed performance level. The Decoder adds that AI performance costs are falling faster than those of any previous technology, and reports a separate MIT estimate of 5-10x annual drops in cost per query after stripping out hardware and competition effects — a slower figure than Epoch's, though the two analyses use different methodologies rather than directly contradicting each other.

• Epoch AI (report by Emberson and Roodman) estimates a given AI performance level became about 47% cheaper per quarter over three years.
• That pace equates to roughly a 13-fold cost drop every year.
• Reaching 75% on GPQA Diamond cost about $0.30 per question with OpenAI o3 in January 2025 versus about $0.0004 with GPT-5.6 Luna in mid-2026 — a roughly 725-fold decline in under 18 months.
• MIT researchers estimate 5-10x annual drops in cost per query after stripping out hardware and competition effects.
• The Decoder reports that AI performance costs are falling faster than those of any previous technology.
• Epoch's ~13x annual figure and MIT's 5-10x annual figure differ, but MIT's estimate deliberately excludes hardware and competition effects.