How fast is AI?
Menu

Updates · Oct 2, 2026

The tracker goes live

The site now syncs itself daily: model releases, capability scores, prices, task horizons, and jobs and productivity data for eight economies.

Until now this site was a snapshot from June. From today it updates itself. Every morning a job pulls new model releases and capability scores from Epoch AI, list prices from providers, task horizons from METR, and labor-market and productivity data from the BLS, the OECD, Eurostat, and the World Bank. Numbers publish automatically when they pass checks; anything editorial waits for review.

The headline from the first sync: GPT-4-level intelligence is now roughly 650 times cheaper per token than when it debuted in March 2023, and 130 models from 14 labs match it. Over roughly the same period, US output per hour rose about 9 percent in total. That gap, a fast clock next to a slow one, is what the weak-link view predicts.

The economy side is no longer US-only. You can switch the slow clock between the US, the UK, Germany, France, Canada, Japan, the EU, and the OECD average. More countries will follow where comparable data exists.

My scenario call for now: reality is still tracking the baseline. Capability is compounding fast, but nothing in the productivity or labor data yet looks like the economy-wide break of the aggressive scenarios.