What the frontier costs
Where model-building actually is, refreshed from the primary sources rather than from anybody's summary of them.
Frontier training compute
Doubling every 5.5 months. Least-squares fit over 34 frontier models released since 2020, computed from Epoch AI's notable-models dataset.
Largest training run
GPT-6 Astra, published 2026-09-03 — the largest training run Epoch AI has been able to document.
Cost of a frontier run
Grok 4, published 2025-07-09. Amortised hardware and energy for a single run, in constant 2023 dollars, as estimated by Epoch AI.
Frontier models with open weights
11% of frontier models released since 2024 ship open weights, per Epoch AI's labelling.
Public models on Hugging Face
3,123,880 public model repositories — the only figure here that moves measurably between one refresh and the next.
On these numbers
Every figure here comes from a machine-readable source — Epoch AI publish the CSV behind their charts, and Hugging Face expose a model index. Nothing on this board is read out of an article.
The compute growth rate is mine, not Epoch’s. It is a least-squares fit over the frontier models in their dataset, and I run it rather than quote them so that the method is visible and the number moves when the data does. It lands at 4.6× a year against the 5× Epoch publish — inside the 4×–6× confidence interval they state, which is the check that the pipeline is honest.
When a source is unreachable the previous value stays, dated and flagged. A stale number that is true beats a fresh one that is invented.