Method, data quality and limits
The race view compares public capability measurements for major Western and Chinese AI labs. It does not claim that a benchmark lead decides who reaches AGI first.
01
Epoch’s Capabilities Index combines scores from more than 50 benchmarks into a general capability scale. We use the highest dated score in each ecosystem as its composite frontier.
ECI is not a percentage, AGI probability, safety score, or arrival forecast.
02
Models are grouped by the country of the organisation reported in the source. Western covers labs in the United States and Europe; Chinese covers mainland Chinese labs.
Cross-border partnerships and multinational teams make any geographic grouping imperfect, so this is an ecosystem lens, not a claim about individual researchers.
03
The roster keeps the three highest-scoring named models per tracked lab and adds the lab’s strongest open-weight model when it is not already represented.
Reasoning-effort variants sharing the same source model name are deduplicated. The highest-scoring representative is kept.
04
We say “open weights”, not “open source”. A model qualifies when the dataset’s accessibility field begins with Open weights.
Restrictions, non-commercial terms and licence obligations are kept in the model table.
05
Rows without a valid score, date or ecosystem identity are excluded. A missing or non-positive result is shown as “no comparable result”, never silently turned into a zero.
22 ECI rows sit outside the scored race because they lack a valid score or a tracked ecosystem classification.
06
Every build pulls Epoch AI’s public benchmark bundle and recomputes the whole site from it. A scheduled daily build keeps it current; if Epoch can’t be reached, the last good snapshot stays up.
This snapshot was built from Epoch data dated October 19, 2026.
Quality notes
Primary references