Epoch Capabilities Index
One number per model, fit by Epoch over the benchmarks that model has results on, with the interval around it. 258 models placed, 116 of them listed in the catalog Vitna prices.
retrieved ·CC-BY-4.0epoch.ai/eci ↗
What the interval says
TOP 24 OF 258Claude Fable 5.2 leads at 173.4, and 8 other models have intervals reaching its own, so the index does not separate them.
- 01Claude Fable 5.2173.4
- 02GPT-6.2 Astra172.8
- 03Gemini 4 Carbon169.7
- 04GPT-6 Astra169.2
- 05Claude Opus 5.5168.7
- 06GPT-6.1 Sol168.6
- 07GPT-6.1 Terra167.7
- 08Claude Fable 5.1167.1
- 09Gemini 4 Hydrogen166.8
- 10Claude Sonnet 5.5166.6
- 11Gemini 4 Argon166.5
- 12Claude Opus 5164.7
- 13Claude Fable 5164.2
- 14GPT-6 Sol163.8
- 15GPT-5.6 Sol163.0
- 16GPT-5.5 Pro162.8
- 17Muse Wonder161.7
- 18Kimi K3.1 Pro161.2
- 19GPT-5.6 Terra160.8
- 20Claude Haiku 5.5160.3
- 21GPT-5.4 Pro160.3
- 22GPT-5.5160.2
- 23Claude Opus 4.8159.9
- 24DeepSeek V4.1 Flash 1017159.4
The rail runs 157.1 to 177.2, the range these 24 rows cover rather than the whole index, because scaled to all 258 they would sit on top of one another. Each bar is the interval Epoch publishes and the mark is the figure; a marked row is one whose interval reaches that of the leader. Epoch states that absolute values are meaningless by themselves and that comparisons between models are what the index is for, that values are not linearly related to benchmark accuracy, and that a highly specialised model may score low while being capable in its own domain.
Vitna fits a scale of its own over the six openly licensed boards, 319 ranked runs over 46 columns, and this is the published index the ordering of that fit is checked against. The two are separate fits over different tables, so their numbers do not compare and only their orderings do.
7 of 258.
Index figures, as Epoch publishes them| # | Model | Organisation | Weights | Released | Results | Interval | Index |
|---|---|---|---|---|---|---|---|
| 18 | Kimi K3.1 Pro | Moonshot | Open weights | 16 | 158.9 to 163.7 | 161.2 | |
| 26 | Kimi K3 | Moonshot | Open weights | 22 | 156.7 to 160.8 | 158.6 | |
| 64 | Kimi K2.6 | Moonshot | Open weights | 15 | 149.9 to 153.4 | 151.6 | |
| 66 | Kimi K2.7 Code | Moonshot | Open weights | 11 | 149.2 to 153.3 | 151.2 | |
| 77 | Kimi K2.5 | Moonshot | Open weights | 14 | 147.3 to 150.3 | 148.8 | |
| 95 | Kimi K2 Thinking | Moonshot | Open weights | 9 | 143.7 to 147.3 | 145.5 | |
| 118 | Kimi K2 (Jul 2025) | Moonshot | Open weights | 6 | 138.4 to 143.9 | 141.3 |
7 of 258 models, highest index first. A model links to its catalog page where the pairing names one, which is true of 116; the rest are models the catalog does not list, and the record Epoch keeps of them stands on its own. Results is how many benchmark results stand behind the figure, and Epoch places nothing on fewer than 4. The two rows with no interval are the anchors the scale is defined by, whose figures are fixed rather than estimated.
Source: the Epoch AI Capabilities Index (epoch.ai/eci), as of 20 Nov 2026. 258 models placed over 2917 benchmark results.
Epoch AI Benchmarking Hub data, by Epoch AI, under CC BY 4.0.
Epoch states that absolute values are meaningless by themselves and that comparisons between models are what the index is for, that values are not linearly related to benchmark accuracy, and that a highly specialised model may score low while being capable in its own domain.
The index is Epoch's fit. Vitna ran none of these evaluations, computed none of these figures and adjusted none of them.
