AI capability
what machines can now do — Epoch AI, METR · dots mark milestones — hover to read, click to open
Frontier model training compute
log₁₀ FLOP (largest model that year)
10^26.7
2025
+9 orders since 2012
The largest training compute (in floating-point operations) of any notable model published that year, on a log scale. AI capability is a story of compute.
2012–2025 · Epoch AI — Notable AI Models database
Notable AI models released
models per year
107
2025
▲ 723% since 2012
Count of models Epoch AI classifies as notable (state-of-the-art, historically significant, or highly cited) with a publication date in that year.
2012–2025 · Epoch AI — Notable AI Models database
AI task time-horizon (50% reliability)
human-minutes of task length
130
2025
×4,300 since 2019
The length of task (in the time it takes a human) that a frontier model can finish with 50% reliability. Doubling roughly every 7 months. Frozen citation read from METR's published figure.
2019–2025 · METR — Measuring AI Ability to Complete Long Tasks