Skip to content
This repository was archived by the owner on Apr 9, 2026. It is now read-only.

Commit dd3529c

Browse files
daniel5ugithub-actions[bot]
authored andcommitted
update data from artificial analysis
1 parent 3e19b80 commit dd3529c

24 files changed

Lines changed: 623 additions & 98 deletions

src/content/benchmarks/artificial_analysis_coding_index.json

Lines changed: 49 additions & 1 deletion
Original file line numberDiff line numberDiff line change
@@ -7,7 +7,7 @@
77
"tags": [
88
"Coding"
99
],
10-
"lastUpdated": "2026-02-12",
10+
"lastUpdated": "2026-02-21",
1111
"metrics": {
1212
"unit": "Index",
1313
"isBetterHigher": true
@@ -19,6 +19,14 @@
1919
"initialWeight": 1000
2020
},
2121
"snapshot": [
22+
{
23+
"modelRef": "google/gemini-3-1-pro-preview",
24+
"score": 55.5
25+
},
26+
{
27+
"modelRef": "anthropic/claude-sonnet-4-6-adaptive",
28+
"score": 50.9
29+
},
2230
{
2331
"modelRef": "openai/gpt-5-2",
2432
"score": 48.7
@@ -43,6 +51,10 @@
4351
"modelRef": "google/gemini-3-pro",
4452
"score": 46.5
4553
},
54+
{
55+
"modelRef": "anthropic/claude-sonnet-4-6",
56+
"score": 46.4
57+
},
4658
{
4759
"modelRef": "openai/gpt-5-1",
4860
"score": 44.7
@@ -59,6 +71,10 @@
5971
"modelRef": "openai/gpt-5-2-codex",
6072
"score": 43.0
6173
},
74+
{
75+
"modelRef": "anthropic/claude-sonnet-4-6-non-reasoning-low-effort",
76+
"score": 43.0
77+
},
6278
{
6379
"modelRef": "anthropic/claude-opus-4-5",
6480
"score": 42.9
@@ -67,6 +83,10 @@
6783
"modelRef": "google/gemini-3-flash-reasoning",
6884
"score": 42.6
6985
},
86+
{
87+
"modelRef": "alibaba/qwen3-5-397b-a17b",
88+
"score": 41.3
89+
},
7090
{
7191
"modelRef": "xai/grok-4",
7292
"score": 40.5
@@ -79,6 +99,10 @@
7999
"modelRef": "google/gemini-3-pro-low",
80100
"score": 39.4
81101
},
102+
{
103+
"modelRef": "zai/glm-5-non-reasoning",
104+
"score": 39.0
105+
},
82106
{
83107
"modelRef": "openai/gpt-5-medium",
84108
"score": 39.0
@@ -103,6 +127,14 @@
103127
"modelRef": "google/gemini-3-flash",
104128
"score": 37.8
105129
},
130+
{
131+
"modelRef": "minimax/minimax-m2-5",
132+
"score": 37.4
133+
},
134+
{
135+
"modelRef": "alibaba/qwen3-5-397b-a17b-non-reasoning",
136+
"score": 37.4
137+
},
106138
{
107139
"modelRef": "deepseek/deepseek-v3-2-reasoning",
108140
"score": 36.7
@@ -1003,6 +1035,10 @@
10031035
"modelRef": "nvidia/nvidia-nemotron-nano-9b-v2",
10041036
"score": 7.5
10051037
},
1038+
{
1039+
"modelRef": "trillionlabs/tri-21b-think-preview",
1040+
"score": 7.4
1041+
},
10061042
{
10071043
"modelRef": "google/gemini-2-5-flash-lite",
10081044
"score": 7.4
@@ -1027,10 +1063,18 @@
10271063
"modelRef": "meta/llama-4-scout",
10281064
"score": 6.7
10291065
},
1066+
{
1067+
"modelRef": "anthropic/claude-3-haiku",
1068+
"score": 6.7
1069+
},
10301070
{
10311071
"modelRef": "alibaba/qwen3-vl-4b-reasoning",
10321072
"score": 6.7
10331073
},
1074+
{
1075+
"modelRef": "trillionlabs/tri-21b-think-v0-5",
1076+
"score": 6.3
1077+
},
10341078
{
10351079
"modelRef": "google/gemma-3-12b",
10361080
"score": 6.3
@@ -1167,6 +1211,10 @@
11671211
"modelRef": "alibaba/qwen3-0-6b-instruct",
11681212
"score": 1.4
11691213
},
1214+
{
1215+
"modelRef": "cohere/tiny-aya-global",
1216+
"score": 1.2
1217+
},
11701218
{
11711219
"modelRef": "ai2/olmo-2-7b",
11721220
"score": 1.2

0 commit comments

Comments
 (0)