346 models tracked today

The AI race, measured

Unbiased, data-driven research on the AI race. Who's actually winning, how fast it's moving, and what it means.

What we track, daily

151 wk

of weekly usage data — who's winning, who's losing

10

data sources, updated every day

3,200+

models — how much they cost to build

104

libraries — which SDKs developers are choosing

158

benchmarks — what actually moved the frontier

1.4T

tokens/day — real enterprise demand, not hype

Published research

Latest: How long can Mythos and Muse Spark work alone? Predicting METR time horizons from IRT benchmark scores across 13 models.

How Long Can Claude Mythos Work Alone?

METR hasn't scored Mythos yet — we predicted it. Regression across 19 models puts it at ~15 hours of autonomous work. If the frontier is accelerating, it could be 30.

Do Benchmark Breakthroughs Actually Matter?

We tested whether SOTA events move stock prices or token consumption. Answer: they shift usage within 2 weeks, but markets don't care. 30+ interactive charts.

Kimi, Composer, and the Build vs Buy Dilemma

Cursor quietly switched from Anthropic to a fine-tuned open-weight model. We traced it in real-time through 6 data feeds before anyone reported it.