Newsbite analysis · AI, Data
The Decoder
AI performance costs are falling faster than those of any previous technology
Epoch AI has identified a striking trend in the AI landscape, with performance costs decreasing approximately 47% each quarter, translating to a staggering 13-fold reduction annually. This rapid decline surpasses that of any previous transformative technology. However, experts caution that these numbers primarily represent market pricing related to benchmark scores, while underlying factors like hardware costs and competitive dynamics play crucial roles. MIT research corroborates some of these findings but offers lower efficiency gain estimates by separating out hardware improvements and competitive pricing factors. As newer models often require more computing power per task, the cost discrepancies underscore the complexity in parsing AI performance metrics.
MonitorConfidence: MediumShort term horizonStory: First report
EvidenceOn record sourcesIndependent evidenceNo quotesArticle type: Analysis
Published 24 Sept 2026, 16:10 UTC5 min read
Lead image — The Decoder
Brief
NewsBite reading:AI performance costs decline rapidly, outperforming past technologies
TL;DR
- Epoch AI shows a 47% quarterly decline in AI performance costs.
- 13x annual cost reduction is noted, the fastest of any prior technology.
- MIT researchers report algorithmic efficiency gains closer to 3x per year.
- Benchmarking issues like 'benchmaxxing' could inflate performance ratings.
- Cost-per-token metrics alone may not indicate real-world performance.
Interesting facts
- An AI scoring 75% on a test in early 2025 cost approximately 30 cents per question.
- A newer model scored the same for just four hundredths of a cent.
- If car prices dropped similarly, a €50,000 car would cost less than €70.