The largest training run ever disclosed

[ Front page ] [ The edge ] [ Finished ] [ About ] [ Links ]

9 x 10^26 operations

Held by GPT-6 Astra. Set 2026-09-03, and has stood for 54 days. Written down here 2026-10-17.

In 1993 this was 2 x 10^13 operations - TD-Gammon, IBM, set 1992-05-01. The figure above is about 45 trillion times that. Where that is checked.

The largest training-compute figure among the 31 models dated 1993 or earlier in Epoch AI's notable-models file, of 1,078 rows, downloaded and sorted in a script: TD-Gammon, published 1992-05-01, at 1.82e13 FLOP, which the file marks Speculative. That is the same file and the same reading rule as the frontier itself, which has never filtered on the Confidence column (2026-09-26). The largest run dated in 1993 itself is Siamese-TDNN at 1.29e13. Printed to one significant figure, as the frontier prints its own readings.

Training a model means doing arithmetic, and the arithmetic can be counted. This is the number of operations used to train the largest model anyone has published a figure for: five hundred septillion, written as five followed by twenty-six zeros. Every computer on earth in 1993, running flat out for a year, would not have come close. The figure is disclosed voluntarily, and several laboratories have stopped disclosing it at all, so this is the largest known run rather than the largest run - a distinction this page keeps rather than hides. If the leaders go quiet, this frontier will stop moving while the thing it measures does not, and that will be worth saying out loud when it happens.

That description was written when this frontier was opened, on 2026-09-07, and is never rewritten. Where it and the number above it disagree, the number is the reading and the description is history.

What counts
The highest training compute figure, in floating-point operations, recorded against any model in Epoch AI's database of notable models. Disclosed figures only, which is not the same as the largest run.
What moves it
A laboratory discloses a larger figure and Epoch AI records it.

Where to check it: https://epoch.ai/data/notable-ai-models

The program reads it from https://epoch.ai/data/notable_ai_models.csv, which is the same figure in a form a script can parse.

Every reading, oldest first

Last checked: 2026-10-27. The source itself last measured this: 2026-09-09.

Previous: Chess puzzles solved   Next: The largest training cluster ever disclosed

[ Front page ] [ The edge ] [ Finished ] [ About ] [ Links ]