Facts · The field · 7 cards
Why the same answer costs 200× less than it used to
Price of the same answer, then and now
These facts come up in the course What AI Costs, where each one is explained step by step.
1
~200×
Matching GPT-4's March 2023 score on MMLU, a broad knowledge test, cost $37.50 per million tokens. In February 2025 Gemini 2.0 Flash did it for $0.18.
$37.50 ÷ $0.18 ≈ 208, over 23 months. The $37.50 blends three input tokens per output token: (3 × $30 + $60) ÷ 4.
2
40× / year
The cheapest model reaching GPT-4 level on GPQA Diamond, a hard science benchmark, got 40 times cheaper per year. Across milestones Epoch found 9 to 900 times a year.
The MMLU fall of 208× in 23 months is 208^(12/23) ≈ 16× a year, slower than the 40× GPQA Diamond rate.
3
About 850 times less in under three years. A small new model soon matches an old large one, so a fixed capability falls in price fast.
4
$38 B
Epoch AI's estimate of the capital cost per gigawatt of IT power, the electricity delivered to the racks. Such a site takes about 2.1 years to build.
At up to about 1,500 W per GPU (chip plus its server, network and cooling), 1 GW holds about 667,000 GPUs. $38 billion ÷ 667,000 ≈ $57,000 per GPU slot, chips and building together.
Source: Epoch AI, Trends, 2026
5
$62 M
Meta reports 30.84 million GPU-hours for its largest Llama 3.1 model. That price excludes salaries, failed experiments and the smaller models.
30.84 million GPU-hours × $2 per hour ≈ $62 million. Meta trained on up to 16,384 H100s; 30.84 million hours is about 78 days of all of them.
6
Wait, really?
By Epoch AI's estimate, only about 30 percent of 2024 compute spending went to inference.
Inference means running the model to answer users. The other seven tenths of the 2024 compute bill went to training and research.
Source: Epoch AI, Trends, 2026
7
Guess first
$0.10
Ten cents per million tokens, about 375 times less in 16 months. A small open model caught up with the old flagship.