Training data and labelling
The Index · September 2026 editionThe companies that produce, label and license the data models are trained on.
Revised edition. First published 2026-09-27; revised as follows:
- 2026-09-28 The Open models table was withdrawn: its dataset contained unsourced and implausible entries and omitted models released in 2026. It will return after a sourced rebuild.
- 2026-09-28 The Frontier labs, Semiconductors and Cloud tables were withdrawn after a review found errors in their data introduced by an automated research process (inflated assessment scores; quarterly revenue recorded as annual; a unit error in data-centre capacity). They will return after the data is restored from sourced values.
How to cite
Cognitive Coefficient Institute (2026). The Cognitive Coefficient Index, September 2026 edition: Training data and labelling. https://alpha.cognitivecoefficient.com/index/editions/2026-09/training-data. Licensed CC BY 4.0.
| # | Name | Segment | SCX · Stack Capacity Index | AIX · AI exposure |
|---|---|---|---|---|
| 01 | Mercor | Expert data & RLHF | 86.1 | 95.0 |
| 02 | Scale AI | Expert data & RLHF | 75.2 | 95.0 |
| 03 | Surge AI | Expert data & RLHF | 74.3 | 100.0 |
| 04 | Handshake (AI) | Expert data & RLHF | 69.5 | 75.0 |
| 05 | Snorkel AI | Evals & RL environments | 54.5 | 95.0 |
| 06 | micro1 | Expert data & RLHF | 53.8 | 100.0 |
| 07 | Innodata | Crowd & annotation services | 42.2 | 85.0 |
| 08 | Invisible Technologies | Expert data & RLHF | 41.2 | 90.0 |
| 09 | Appen | Crowd & annotation services | 35.7 | 90.0 |
| 10 | Labelbox | Labelling platforms | 32.9 | 95.0 |
Stack Capacity Index: components and weights in this edition
Licensed CC BY 4.0: reuse freely, credit “Cognitive Coefficient Institute”. Scores are model outputs, not predictions. Terms.