Models · hardware · harnesses · data as of 24 Sep 2026

Artificial Acceleration
How fast AI is moving: the models, the machines, the agents.

Every frontier lab published calls to pace or throttle AI development in 2026, while shipping faster than ever. Nine labs over four years, every release dated, measured two ways: how many models shipped, and how much smarter they got. Then what it costs to run them, and the agents that put them to work.

Appearance

Intelligence is models × hardware × harnesses: what the labs release, what it costs to run, and the agents that put it to work. This site tracks all three.

01Models31 releases in 90 daysfrom 9 labs. Smartest on the Intelligence Index: Gemini 4 Carbon (Google, 60.0).releases per month, two yearsThe release record →02Hardware$3.80 an hour to rent an H100+29% on a year ago. The cheapest RTX 5090 in stock in the US costs $7,550; see 17 countries on the map.H100 median $/GPU-hour, 12 monthsPrices around the world →03Harnesses59M agent downloads last week9× a year ago, across 37 agents. Best agent on Terminal-Bench 4.0: Claude Code + Opus 5.5, 64.8%.weekly npm + PyPI downloads, 12 monthsThe agent tracker →
Window
Chart mode
Labs · click to isolate
ringed = a notable release we write up small = everything else in the record line = that lab’s release range ❝a lab said it would slow down — hover to see the next 30 days

Releases over time

Show these releases as a table