What happened
Epoch AI published a quantitative analysis quantifying how fast AI capability has become cheaper, concluding 'AI has gotten cheaper more quickly than any other transformative technology in history.' The report estimates 'the cost of achieving a given level of AI performance has fallen about 47% per quarter since 2023, or 13× per year' — roughly four times faster than DNA sequencing, six times faster than compute, 18 times faster than lithium batteries, and 54 times faster than the historical price decline of electricity. Its headline example: OpenAI's o3 achieved ~75% on GPQA Diamond at an estimated $0.30 per question in early 2025, while GPT-5.6 Luna matched that score for $0.0004 per question — a 725-fold drop in under 18 months. The analysis builds a new dataset of more than 222 AI models across up to 11 benchmarks, using budget-constrained cost-performance curves, and cautions on benchmark contamination, the short timeframe, and that real users rarely switch to the absolute cheapest model.
Why it matters
The falling 'price of thought' is the economic signal behind both AI-enabled business-model shifts and the democratisation of capability, and it directly informs infrastructure and vendor-cost assumptions in AI business cases.
Action needed
Use the cost-curve data to refresh assumptions in AI ROI and scaling models; expect capability previously reserved for frontier vendors to become commodity-priced, and scenario-plan vendor pricing accordingly.