Monthly counts of models in Epoch AI's compute-defined frontier subset, grouped by lab from January 2023. The series shows which developers appear at the training-compute frontier and how release activity is distributed, but it does not count every AI model; source coverage ends with a July 2025 publication.
Frontier Model Releases per Month by Lab
OpenAI
Google
Anthropic
xAI
Meta AI
NVIDIA
Z.ai
Other
SOURCE: Epoch AI — Frontier AI Models
As of 2026-07-31, OpenAI records 0 models.
Pro API coming soon
Methodology
Source inclusion is not re-created or broadened by The Latent. A row qualifies only because it appears in Epoch AI's Frontier AI Models CSV and its Frontier model field is true. Epoch's exact definition is: models that were in the top 10 by training compute as of their release date.
Filter Publication date from 2023-01-01 through 2026-07-31, the latest complete month as of 2026-08-27. Count each unique Model row once in the calendar month of Publication date. The timestamp stored for every bucket is that calendar month's final day.
Epoch defines Publication date as the publication, announcement, or release date. When only year and month are known, Epoch fills day 15; when only the year is known, it fills July 1. Monthly placement therefore inherits any date imputation in the source.
The readable lab set is fixed: OpenAI, Google, Anthropic, xAI, Meta AI, NVIDIA, and Z.ai. These are the normalized lab labels with at least two records in the selected snapshot and are held fixed for future updates. Google combines only the exact Epoch organization labels Google and Google DeepMind. Every other organization label is assigned to Other, making the sum across buckets exactly equal to the selected source rows.
Every month-lab combination is emitted. A zero means the downloaded Epoch frontier subset has no row in that bucket; it does not mean the lab released no models. Epoch describes its broader model list as non-exhaustive, and frontier classification depends on available or estimated training-compute information.
This is a derived count. Model names, organization labels, dates, and frontier inclusion remain attributable to Epoch AI; no attempt is made to infer omitted models, split one source row into model variants, or retroactively reconstruct the frontier.
Frequently asked questions
What does Epoch AI mean by a frontier model?
Epoch defines frontier models as models that were in the top 10 by training compute as of their release date. This chart uses that source flag exactly.
Does this chart count every AI model release?
No. It counts records in Epoch AI's Frontier AI Models subset. The source database is non-exhaustive, and inclusion requires enough training-compute information for Epoch's frontier classification.
Why is frontier release activity relevant to the AI industry?
It shows which labs appear in a source-defined, training-compute frontier and how frequently qualifying models arrive. That can contextualize competitive release cadence and concentration, but it does not measure model quality, adoption, revenue, or overall market share.
Why are there zeroes after July 2025?
The downloaded snapshot's latest Publication date is 2025-07-09. Zeroes through 2026-07-31 preserve the requested complete monthly axis but mean only that the snapshot contains no qualifying rows in those buckets.
How are labs grouped?
The chart keeps seven fixed lab series. Google and Google DeepMind source labels are combined as Google; exact OpenAI, Anthropic, xAI, Meta AI, NVIDIA, and Z.ai (Zhipu AI) labels retain their own series. All remaining labels are counted in Other.
Can a model's month or frontier status change later?
Yes. Epoch may revise publication dates, organization labels, training-compute estimates, or frontier flags as evidence changes. The derived chart should therefore be regenerated from the current authoritative download.