AI Model Tracker #

The frontier over time #

Each line is one lab's strongest model on that date, across 231 published boards. It opens on the 6 labs that have actually held the top slot, plus Mistral. The other 6 are greyed out in the legend: click any name to add it, double-click to isolate one.

Who is being used #

Share of the 2,125,549 head-to-head comparisons recorded between 2026-05-01 and 2026-07-30. This measures what people chose to run a model against, not installed base or revenue, and the gap between it and the ranking above is the interesting part: labs can be in front on quality and behind on reach, or the reverse. Bars are coloured by where the lab is domiciled.

CSV

The labs and what they are worth #

Last reported financing event per lab, not a mark. Private valuations are whatever the most recent round or share sale implied, on the date it was reported, and several of the largest have never been tested by a public market. Everything is shown in euros at today's rate so the column is comparable; where that differs from the announced figure, the original is printed underneath. Where the lab sits inside a listed company, the ticker points at the parent instead. Every figure links to its source.

Rank orders the labs by the rating of their strongest model. The smaller figure beneath is where that model sits among all 383 models on the board, which is a different question: a lab shipping several strong checkpoints occupies several of the top slots and pushes everyone else down a list they are not really behind on. Right now Anthropic holds model ranks 1 to 7 with 7 of its own, so Google DeepMind is second among labs but only 8th among models.

Round status: closed announced as completed by the company or a tier-1 outlet; reported reported as in progress or in talks, not confirmed closed; secondary price set by a share sale between investors or employees, not a primary raise

The board #

Ratings come from blind pairwise comparisons, so they measure what people prefer rather than what a benchmark scores. The interval is the published confidence band: two models whose bands overlap are not separated by the data. Click any column header to sort. Pick a category to see where the order changes.

Who is tied to whom #

The frontier is not a set of independent competitors. Chip suppliers sit on the cap tables of their own customers, cloud providers own stakes in the labs renting their capacity, and two of the largest listed names have better exposure through a shareholding than through the models they ship themselves.