The open era moved — and took the evidence with it

Epoch AI · Notable AI Models · CC BY 4.0 · 1950–2026 · 1,078 models

The first dashboard here that is not frozen. A connection-backed Postgres dataset on a daily refresh that rebuilds its wasm extract. It reads as of 2026-09-30, the newest publication date in the rows, and that date moves on its own.

A dashboard titled "The open era moved — and took the evidence with it". Three statistic cards read "the open share of released models" (100.0% in 2019 to 35.0%), "it did not fall everywhere — since 2024" (China 67.8% against the USA’s 21.2%) and "the largest training run anyone has confirmed" (3.8×10²⁵ FLOP, Llama 3.1-405B, open weights). Below them a line chart of the open-weights share by year falling through a dashed "half" line in 2024, with grey bars behind it for the number of models released each year; then a log-scale scatter of every training run coloured by whether the weights are open and faded by how confident the figure is, a list of the eight closed models believed to be bigger than anything confidently measured, country and confidence bar rows, and a median-training-cost chart whose coverage bars collapse after 2023.

A screenshot of the published dashboard, captured from the same URL anybody can open.

Data lane: Query lane
Per view: 78.1 KB fetched, 1 extract
Data age: as of 2026-09-30
Reachability: opens with no account

Measured against the live host, with no credential, on 2026-10-08.

The question it answers

Epoch AI tracks every notable model since 1950. Over that record, did openness follow capability?

The data, and over what period

Training compute, training cost in 2023 dollars, accessibility, country, and Epoch’s own confidence in each figure.

How it gets its rows

One query, in your browser, with no key.

What comes back

1,078 rows × 19 columns.

What was hard

The denominator changes the finding. Epoch marks 287 models “Unreleased”, mostly academic work that never shipped. Counted as closed, the open share peaks at 56% and drifts to 33%. Counted out, as here, it goes from 100% in 2019 to 35%.

What happened after the build

Refreshed daily.

The dashboard itself

Open the live dashboard

Still image

The live dashboard loads from the viewer host. Nothing is fetched from it until you ask, so this page does not touch that origin unless you want it to.

https://share.vibedasher.com/eu/vibedasher/notable-ai-models-the-open-era-moved?exec=wasm

What loads here is the published dashboard, from the same address anybody else would open — not a copy kept by hand. The still above it is a screenshot of that same address.

Published by Vibedasher