orcacompass 0.0.0-stage → 0.1.0
This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
- package/LICENSE +202 -0
- package/NOTICE +51 -0
- package/README.md +171 -3
- package/data/modelindex-snapshot.json +1537 -0
- package/data/models-snapshot.json +1356 -0
- package/dist/cli.js +906 -0
- package/dist/config.js +272 -0
- package/dist/install.js +56 -0
- package/dist/paths.js +30 -0
- package/dist/placements/agents/antigravity.js +87 -0
- package/dist/placements/agents/claude.js +72 -0
- package/dist/placements/agents/codex.js +66 -0
- package/dist/placements/agents/cursor.js +102 -0
- package/dist/placements/agents/deps.js +19 -0
- package/dist/placements/agents/grok.js +92 -0
- package/dist/placements/agents/list-cache.js +85 -0
- package/dist/placements/agents/muse.js +7 -0
- package/dist/placements/agents/opencode.js +77 -0
- package/dist/placements/agents/orca-catalog.js +34 -0
- package/dist/placements/benchmark.js +70 -0
- package/dist/placements/build.js +94 -0
- package/dist/placements/score.js +70 -0
- package/dist/placements/sources/artificialanalysis.js +133 -0
- package/dist/placements/sources/modelindex.js +84 -0
- package/dist/placements/sources/modelsdev.js +113 -0
- package/dist/placements/table.js +94 -0
- package/dist/placements/tiers.js +136 -0
- package/dist/routing/effort.js +31 -0
- package/dist/routing/environment.js +51 -0
- package/dist/routing/orca-settings.js +111 -0
- package/dist/routing/orca-usage.js +85 -0
- package/dist/routing/pace.js +52 -0
- package/dist/routing/rejection.js +58 -0
- package/dist/routing/rejections.js +18 -0
- package/dist/routing/router.js +552 -0
- package/dist/setup.js +358 -0
- package/dist/system/run.js +128 -0
- package/dist/types.js +8 -0
- package/dist/verify.js +106 -0
- package/dist/version.js +3 -0
- package/package.json +60 -6
- package/skills/orcacompass/SKILL.md +278 -0
|
@@ -0,0 +1,1537 @@
|
|
|
1
|
+
{
|
|
2
|
+
"fetchedAt": 1793354566936,
|
|
3
|
+
"ratings": [
|
|
4
|
+
{
|
|
5
|
+
"id": "gpt-4-jun-2023",
|
|
6
|
+
"name": "GPT-4 (Jun 2023)",
|
|
7
|
+
"date": "2023-06-13",
|
|
8
|
+
"score": 101,
|
|
9
|
+
"out": 60,
|
|
10
|
+
"orId": "openai/gpt-4",
|
|
11
|
+
"open": false
|
|
12
|
+
},
|
|
13
|
+
{
|
|
14
|
+
"id": "mistral-large",
|
|
15
|
+
"name": "Mistral Large",
|
|
16
|
+
"date": "2024-02-26",
|
|
17
|
+
"score": 100.7,
|
|
18
|
+
"out": 6,
|
|
19
|
+
"orId": "mistralai/mistral-large",
|
|
20
|
+
"open": false
|
|
21
|
+
},
|
|
22
|
+
{
|
|
23
|
+
"id": "mixtral-8x22b",
|
|
24
|
+
"name": "Mixtral 8x22B",
|
|
25
|
+
"date": "2024-04-10",
|
|
26
|
+
"score": 99.3,
|
|
27
|
+
"out": 6,
|
|
28
|
+
"orId": "mistralai/mixtral-8x22b-instruct",
|
|
29
|
+
"hf": "mistralai/Mixtral-8x22B-Instruct-v0.1",
|
|
30
|
+
"open": true
|
|
31
|
+
},
|
|
32
|
+
{
|
|
33
|
+
"id": "gpt-4o",
|
|
34
|
+
"name": "GPT-4o",
|
|
35
|
+
"date": "2024-05-13",
|
|
36
|
+
"score": 111.2,
|
|
37
|
+
"out": 10,
|
|
38
|
+
"orId": "openai/gpt-4o",
|
|
39
|
+
"open": false
|
|
40
|
+
},
|
|
41
|
+
{
|
|
42
|
+
"id": "gemma-2",
|
|
43
|
+
"name": "Gemma 2",
|
|
44
|
+
"date": "2024-06-27",
|
|
45
|
+
"score": 100.5,
|
|
46
|
+
"out": 0.65,
|
|
47
|
+
"orId": "google/gemma-2-27b-it",
|
|
48
|
+
"hf": "google/gemma-2-27b-it",
|
|
49
|
+
"open": true
|
|
50
|
+
},
|
|
51
|
+
{
|
|
52
|
+
"id": "gpt-4o-mini",
|
|
53
|
+
"name": "GPT-4o mini",
|
|
54
|
+
"date": "2024-07-18",
|
|
55
|
+
"score": 104.8,
|
|
56
|
+
"out": 0.6,
|
|
57
|
+
"orId": "openai/gpt-4o-mini",
|
|
58
|
+
"open": false
|
|
59
|
+
},
|
|
60
|
+
{
|
|
61
|
+
"id": "mistral-nemo",
|
|
62
|
+
"name": "Mistral NeMo",
|
|
63
|
+
"date": "2024-07-18",
|
|
64
|
+
"score": 91.8,
|
|
65
|
+
"out": 0.03,
|
|
66
|
+
"orId": "mistralai/mistral-nemo",
|
|
67
|
+
"hf": "mistralai/Mistral-Nemo-Instruct-2407",
|
|
68
|
+
"open": true
|
|
69
|
+
},
|
|
70
|
+
{
|
|
71
|
+
"id": "llama-3-1-70b",
|
|
72
|
+
"name": "Llama 3.1-70B",
|
|
73
|
+
"date": "2024-07-23",
|
|
74
|
+
"score": 105.6,
|
|
75
|
+
"out": 0.4,
|
|
76
|
+
"orId": "meta-llama/llama-3.1-70b-instruct",
|
|
77
|
+
"hf": "meta-llama/Meta-Llama-3.1-70B-Instruct",
|
|
78
|
+
"open": true
|
|
79
|
+
},
|
|
80
|
+
{
|
|
81
|
+
"id": "llama-3-1-8b",
|
|
82
|
+
"name": "Llama 3.1-8B",
|
|
83
|
+
"date": "2024-07-23",
|
|
84
|
+
"score": 91.2,
|
|
85
|
+
"out": 0.08,
|
|
86
|
+
"orId": "meta-llama/llama-3.1-8b-instruct",
|
|
87
|
+
"hf": "meta-llama/Meta-Llama-3.1-8B-Instruct",
|
|
88
|
+
"open": true
|
|
89
|
+
},
|
|
90
|
+
{
|
|
91
|
+
"id": "gpt-4o-aug-2024",
|
|
92
|
+
"name": "GPT-4o (Aug 2024)",
|
|
93
|
+
"date": "2024-08-06",
|
|
94
|
+
"score": 111.8,
|
|
95
|
+
"out": 10,
|
|
96
|
+
"orId": "openai/gpt-4o-2024-08-06",
|
|
97
|
+
"open": false
|
|
98
|
+
},
|
|
99
|
+
{
|
|
100
|
+
"id": "qwen2-5",
|
|
101
|
+
"name": "Qwen2.5",
|
|
102
|
+
"date": "2024-09-19",
|
|
103
|
+
"score": 113,
|
|
104
|
+
"out": 0.4,
|
|
105
|
+
"orId": "qwen/qwen-2.5-72b-instruct",
|
|
106
|
+
"hf": "Qwen/Qwen2.5-72B-Instruct",
|
|
107
|
+
"open": true
|
|
108
|
+
},
|
|
109
|
+
{
|
|
110
|
+
"id": "qwen2-5-7b",
|
|
111
|
+
"name": "Qwen2.5-7B",
|
|
112
|
+
"date": "2024-09-19",
|
|
113
|
+
"score": 95.2,
|
|
114
|
+
"out": 0.2,
|
|
115
|
+
"orId": "qwen/qwen-2.5-7b-instruct",
|
|
116
|
+
"hf": "Qwen/Qwen2.5-7B-Instruct",
|
|
117
|
+
"open": true
|
|
118
|
+
},
|
|
119
|
+
{
|
|
120
|
+
"id": "llama-3-2-1b",
|
|
121
|
+
"name": "Llama 3.2 1B",
|
|
122
|
+
"date": "2024-09-24",
|
|
123
|
+
"score": 80.3,
|
|
124
|
+
"out": 0.201,
|
|
125
|
+
"orId": "meta-llama/llama-3.2-1b-instruct",
|
|
126
|
+
"hf": "meta-llama/Llama-3.2-1B-Instruct",
|
|
127
|
+
"open": true
|
|
128
|
+
},
|
|
129
|
+
{
|
|
130
|
+
"id": "llama-3-2-3b",
|
|
131
|
+
"name": "Llama 3.2 3B",
|
|
132
|
+
"date": "2024-09-24",
|
|
133
|
+
"score": 93.4,
|
|
134
|
+
"out": 0.33,
|
|
135
|
+
"orId": "meta-llama/llama-3.2-3b-instruct",
|
|
136
|
+
"hf": "meta-llama/Llama-3.2-3B-Instruct",
|
|
137
|
+
"open": true
|
|
138
|
+
},
|
|
139
|
+
{
|
|
140
|
+
"id": "qwen2-5-coder-32b",
|
|
141
|
+
"name": "Qwen2.5-Coder-32B",
|
|
142
|
+
"date": "2024-11-12",
|
|
143
|
+
"score": 111.2,
|
|
144
|
+
"out": 1,
|
|
145
|
+
"orId": "qwen/qwen-2.5-coder-32b-instruct",
|
|
146
|
+
"hf": "Qwen/Qwen2.5-Coder-32B-Instruct",
|
|
147
|
+
"open": true
|
|
148
|
+
},
|
|
149
|
+
{
|
|
150
|
+
"id": "gpt-4o-nov-2024",
|
|
151
|
+
"name": "GPT-4o (Nov 2024)",
|
|
152
|
+
"date": "2024-11-20",
|
|
153
|
+
"score": 110.7,
|
|
154
|
+
"out": 10,
|
|
155
|
+
"orId": "openai/gpt-4o-2024-11-20",
|
|
156
|
+
"open": false
|
|
157
|
+
},
|
|
158
|
+
{
|
|
159
|
+
"id": "llama-3-3-70b",
|
|
160
|
+
"name": "Llama 3.3 70B",
|
|
161
|
+
"date": "2024-12-06",
|
|
162
|
+
"score": 107.8,
|
|
163
|
+
"out": 0.32,
|
|
164
|
+
"orId": "meta-llama/llama-3.3-70b-instruct",
|
|
165
|
+
"hf": "meta-llama/Llama-3.3-70B-Instruct",
|
|
166
|
+
"open": true
|
|
167
|
+
},
|
|
168
|
+
{
|
|
169
|
+
"id": "phi-4",
|
|
170
|
+
"name": "Phi-4",
|
|
171
|
+
"date": "2024-12-12",
|
|
172
|
+
"score": 114.7,
|
|
173
|
+
"out": 0.14,
|
|
174
|
+
"orId": "microsoft/phi-4",
|
|
175
|
+
"hf": "microsoft/phi-4",
|
|
176
|
+
"open": true
|
|
177
|
+
},
|
|
178
|
+
{
|
|
179
|
+
"id": "deepseek-v3",
|
|
180
|
+
"name": "DeepSeek-V3",
|
|
181
|
+
"date": "2024-12-26",
|
|
182
|
+
"score": 118.8,
|
|
183
|
+
"out": 1.0287,
|
|
184
|
+
"orId": "deepseek/deepseek-chat",
|
|
185
|
+
"hf": "deepseek-ai/DeepSeek-V3",
|
|
186
|
+
"open": true
|
|
187
|
+
},
|
|
188
|
+
{
|
|
189
|
+
"id": "qwen-plus-jan-2025",
|
|
190
|
+
"name": "Qwen Plus (Jan 2025)",
|
|
191
|
+
"date": "2025-01-25",
|
|
192
|
+
"score": 115.1,
|
|
193
|
+
"out": 0.78,
|
|
194
|
+
"orId": "qwen/qwen-plus",
|
|
195
|
+
"open": false
|
|
196
|
+
},
|
|
197
|
+
{
|
|
198
|
+
"id": "mistral-small-3",
|
|
199
|
+
"name": "Mistral Small 3",
|
|
200
|
+
"date": "2025-01-30",
|
|
201
|
+
"score": 108.3,
|
|
202
|
+
"out": 0.08,
|
|
203
|
+
"orId": "mistralai/mistral-small-24b-instruct-2501",
|
|
204
|
+
"hf": "mistralai/Mistral-Small-24B-Instruct-2501",
|
|
205
|
+
"open": true
|
|
206
|
+
},
|
|
207
|
+
{
|
|
208
|
+
"id": "gemma-3",
|
|
209
|
+
"name": "Gemma 3",
|
|
210
|
+
"date": "2025-03-12",
|
|
211
|
+
"score": 114.3,
|
|
212
|
+
"out": 0.35,
|
|
213
|
+
"orId": "google/gemma-3-27b-it",
|
|
214
|
+
"hf": "google/gemma-3-27b-it",
|
|
215
|
+
"open": true
|
|
216
|
+
},
|
|
217
|
+
{
|
|
218
|
+
"id": "gemma-3-12b",
|
|
219
|
+
"name": "Gemma 3 12B",
|
|
220
|
+
"date": "2025-03-12",
|
|
221
|
+
"score": 106.7,
|
|
222
|
+
"out": 0.15,
|
|
223
|
+
"orId": "google/gemma-3-12b-it",
|
|
224
|
+
"hf": "google/gemma-3-12b-it",
|
|
225
|
+
"open": true
|
|
226
|
+
},
|
|
227
|
+
{
|
|
228
|
+
"id": "gemma-3-4b",
|
|
229
|
+
"name": "Gemma 3 4B",
|
|
230
|
+
"date": "2025-03-12",
|
|
231
|
+
"score": 88.3,
|
|
232
|
+
"out": 0.1,
|
|
233
|
+
"orId": "google/gemma-3-4b-it",
|
|
234
|
+
"hf": "google/gemma-3-4b-it",
|
|
235
|
+
"open": true
|
|
236
|
+
},
|
|
237
|
+
{
|
|
238
|
+
"id": "mistral-small-3-1",
|
|
239
|
+
"name": "Mistral Small 3.1",
|
|
240
|
+
"date": "2025-03-17",
|
|
241
|
+
"score": 108.2,
|
|
242
|
+
"out": 0.555,
|
|
243
|
+
"orId": "mistralai/mistral-small-3.1-24b-instruct",
|
|
244
|
+
"hf": "mistralai/Mistral-Small-3.1-24B-Instruct-2503",
|
|
245
|
+
"open": true
|
|
246
|
+
},
|
|
247
|
+
{
|
|
248
|
+
"id": "deepseek-v3-mar-2025",
|
|
249
|
+
"name": "DeepSeek-V3 (Mar 2025)",
|
|
250
|
+
"date": "2025-03-24",
|
|
251
|
+
"score": 124.8,
|
|
252
|
+
"out": 1.14,
|
|
253
|
+
"orId": "deepseek/deepseek-chat-v3-0324",
|
|
254
|
+
"hf": "deepseek-ai/DeepSeek-V3-0324",
|
|
255
|
+
"open": true
|
|
256
|
+
},
|
|
257
|
+
{
|
|
258
|
+
"id": "llama-4-maverick",
|
|
259
|
+
"name": "Llama 4 Maverick",
|
|
260
|
+
"date": "2025-04-05",
|
|
261
|
+
"score": 117.2,
|
|
262
|
+
"out": 0.6525,
|
|
263
|
+
"orId": "meta-llama/llama-4-maverick",
|
|
264
|
+
"hf": "meta-llama/Llama-4-Maverick-17B-128E-Instruct",
|
|
265
|
+
"open": true
|
|
266
|
+
},
|
|
267
|
+
{
|
|
268
|
+
"id": "llama-4-scout",
|
|
269
|
+
"name": "Llama 4 Scout",
|
|
270
|
+
"date": "2025-04-05",
|
|
271
|
+
"score": 112.8,
|
|
272
|
+
"out": 0.3,
|
|
273
|
+
"orId": "meta-llama/llama-4-scout",
|
|
274
|
+
"hf": "meta-llama/Llama-4-Scout-17B-16E-Instruct",
|
|
275
|
+
"open": true
|
|
276
|
+
},
|
|
277
|
+
{
|
|
278
|
+
"id": "gpt-4-1",
|
|
279
|
+
"name": "GPT-4.1",
|
|
280
|
+
"date": "2025-04-14",
|
|
281
|
+
"score": 126.9,
|
|
282
|
+
"out": 8,
|
|
283
|
+
"orId": "openai/gpt-4.1",
|
|
284
|
+
"open": false
|
|
285
|
+
},
|
|
286
|
+
{
|
|
287
|
+
"id": "gpt-4-1-mini",
|
|
288
|
+
"name": "GPT-4.1 mini",
|
|
289
|
+
"date": "2025-04-14",
|
|
290
|
+
"score": 121.3,
|
|
291
|
+
"out": 1.6,
|
|
292
|
+
"orId": "openai/gpt-4.1-mini",
|
|
293
|
+
"open": false
|
|
294
|
+
},
|
|
295
|
+
{
|
|
296
|
+
"id": "gpt-4-1-nano",
|
|
297
|
+
"name": "GPT-4.1 nano",
|
|
298
|
+
"date": "2025-04-14",
|
|
299
|
+
"score": 110.7,
|
|
300
|
+
"out": 0.4,
|
|
301
|
+
"orId": "openai/gpt-4.1-nano",
|
|
302
|
+
"open": false
|
|
303
|
+
},
|
|
304
|
+
{
|
|
305
|
+
"id": "o3-2025-04",
|
|
306
|
+
"name": "o3",
|
|
307
|
+
"date": "2025-04-16",
|
|
308
|
+
"score": 149.2,
|
|
309
|
+
"out": 8,
|
|
310
|
+
"orId": "openai/o3",
|
|
311
|
+
"open": false
|
|
312
|
+
},
|
|
313
|
+
{
|
|
314
|
+
"id": "qwen3-14b",
|
|
315
|
+
"name": "Qwen3-14B",
|
|
316
|
+
"date": "2025-04-28",
|
|
317
|
+
"score": 130.6,
|
|
318
|
+
"out": 0.24,
|
|
319
|
+
"orId": "qwen/qwen3-14b",
|
|
320
|
+
"hf": "Qwen/Qwen3-14B",
|
|
321
|
+
"open": true
|
|
322
|
+
},
|
|
323
|
+
{
|
|
324
|
+
"id": "qwen3-30b-a3b",
|
|
325
|
+
"name": "Qwen3-30B-A3B",
|
|
326
|
+
"date": "2025-04-28",
|
|
327
|
+
"score": 126.6,
|
|
328
|
+
"out": 0.5,
|
|
329
|
+
"orId": "qwen/qwen3-30b-a3b",
|
|
330
|
+
"hf": "Qwen/Qwen3-30B-A3B",
|
|
331
|
+
"open": true
|
|
332
|
+
},
|
|
333
|
+
{
|
|
334
|
+
"id": "qwen3-32b",
|
|
335
|
+
"name": "Qwen3-32B",
|
|
336
|
+
"date": "2025-04-28",
|
|
337
|
+
"score": 130.6,
|
|
338
|
+
"out": 0.28,
|
|
339
|
+
"orId": "qwen/qwen3-32b",
|
|
340
|
+
"hf": "Qwen/Qwen3-32B",
|
|
341
|
+
"open": true
|
|
342
|
+
},
|
|
343
|
+
{
|
|
344
|
+
"id": "qwen3",
|
|
345
|
+
"name": "Qwen3",
|
|
346
|
+
"date": "2025-04-29",
|
|
347
|
+
"score": 134.4,
|
|
348
|
+
"out": 0.5,
|
|
349
|
+
"orId": "qwen/qwen3-30b-a3b",
|
|
350
|
+
"hf": "Qwen/Qwen3-30B-A3B",
|
|
351
|
+
"open": true
|
|
352
|
+
},
|
|
353
|
+
{
|
|
354
|
+
"id": "mistral-medium-3",
|
|
355
|
+
"name": "Mistral Medium 3",
|
|
356
|
+
"date": "2025-05-07",
|
|
357
|
+
"score": 121.4,
|
|
358
|
+
"out": 2,
|
|
359
|
+
"orId": "mistralai/mistral-medium-3",
|
|
360
|
+
"open": false
|
|
361
|
+
},
|
|
362
|
+
{
|
|
363
|
+
"id": "gemini-2-5-pro-jun-2025",
|
|
364
|
+
"name": "Gemini 2.5 Pro (Jun 2025)",
|
|
365
|
+
"date": "2025-06-05",
|
|
366
|
+
"score": 146.3,
|
|
367
|
+
"out": 10,
|
|
368
|
+
"orId": "google/gemini-2.5-pro",
|
|
369
|
+
"open": false
|
|
370
|
+
},
|
|
371
|
+
{
|
|
372
|
+
"id": "gemini-2-5-flash-jun-2025",
|
|
373
|
+
"name": "Gemini 2.5 Flash (Jun 2025)",
|
|
374
|
+
"date": "2025-06-17",
|
|
375
|
+
"score": 132.4,
|
|
376
|
+
"out": 2.5,
|
|
377
|
+
"orId": "google/gemini-2.5-flash",
|
|
378
|
+
"open": false
|
|
379
|
+
},
|
|
380
|
+
{
|
|
381
|
+
"id": "gemini-2-5-flash-lite",
|
|
382
|
+
"name": "Gemini 2.5 Flash-Lite",
|
|
383
|
+
"date": "2025-06-17",
|
|
384
|
+
"score": 121.4,
|
|
385
|
+
"out": 0.4,
|
|
386
|
+
"orId": "google/gemini-2.5-flash-lite",
|
|
387
|
+
"open": false
|
|
388
|
+
},
|
|
389
|
+
{
|
|
390
|
+
"id": "mistral-small-3-2",
|
|
391
|
+
"name": "Mistral Small 3.2",
|
|
392
|
+
"date": "2025-06-20",
|
|
393
|
+
"score": 118.5,
|
|
394
|
+
"out": 0.2,
|
|
395
|
+
"orId": "mistralai/mistral-small-3.2-24b-instruct",
|
|
396
|
+
"hf": "mistralai/Mistral-Small-3.2-24B-Instruct-2506",
|
|
397
|
+
"open": true
|
|
398
|
+
},
|
|
399
|
+
{
|
|
400
|
+
"id": "kimi-k2",
|
|
401
|
+
"name": "Kimi K2",
|
|
402
|
+
"date": "2025-07-11",
|
|
403
|
+
"score": 132.5,
|
|
404
|
+
"out": 2.3,
|
|
405
|
+
"orId": "moonshotai/kimi-k2",
|
|
406
|
+
"hf": "moonshotai/Kimi-K2-Instruct",
|
|
407
|
+
"open": true
|
|
408
|
+
},
|
|
409
|
+
{
|
|
410
|
+
"id": "glm-4-5-air",
|
|
411
|
+
"name": "GLM-4.5-Air",
|
|
412
|
+
"date": "2025-07-20",
|
|
413
|
+
"score": 120.9,
|
|
414
|
+
"out": 0.85,
|
|
415
|
+
"orId": "z-ai/glm-4.5-air",
|
|
416
|
+
"hf": "zai-org/GLM-4.5-Air",
|
|
417
|
+
"open": true
|
|
418
|
+
},
|
|
419
|
+
{
|
|
420
|
+
"id": "qwen3-coder",
|
|
421
|
+
"name": "Qwen3-Coder",
|
|
422
|
+
"date": "2025-07-22",
|
|
423
|
+
"score": 134.9,
|
|
424
|
+
"out": 1,
|
|
425
|
+
"orId": "qwen/qwen3-coder",
|
|
426
|
+
"hf": "Qwen/Qwen3-Coder-480B-A35B-Instruct",
|
|
427
|
+
"open": true
|
|
428
|
+
},
|
|
429
|
+
{
|
|
430
|
+
"id": "qwen3-235b-a22b-instruct-jul-2025",
|
|
431
|
+
"name": "Qwen3-235B-A22B-Instruct (Jul 2025)",
|
|
432
|
+
"date": "2025-07-25",
|
|
433
|
+
"score": 131.7,
|
|
434
|
+
"out": 0.55,
|
|
435
|
+
"orId": "qwen/qwen3-235b-a22b-2507",
|
|
436
|
+
"hf": "Qwen/Qwen3-235B-A22B-Instruct-2507",
|
|
437
|
+
"open": true
|
|
438
|
+
},
|
|
439
|
+
{
|
|
440
|
+
"id": "qwen3-235b-a22b-thinking-jul-2025",
|
|
441
|
+
"name": "Qwen3-235B-A22B-Thinking (Jul 2025)",
|
|
442
|
+
"date": "2025-07-25",
|
|
443
|
+
"score": 140.1,
|
|
444
|
+
"out": 3.5,
|
|
445
|
+
"orId": "qwen/qwen3-235b-a22b-thinking-2507",
|
|
446
|
+
"hf": "Qwen/Qwen3-235B-A22B-Thinking-2507",
|
|
447
|
+
"open": true
|
|
448
|
+
},
|
|
449
|
+
{
|
|
450
|
+
"id": "glm-4-5",
|
|
451
|
+
"name": "GLM-4.5",
|
|
452
|
+
"date": "2025-07-28",
|
|
453
|
+
"score": 133.4,
|
|
454
|
+
"out": 2.2,
|
|
455
|
+
"orId": "z-ai/glm-4.5",
|
|
456
|
+
"hf": "zai-org/GLM-4.5",
|
|
457
|
+
"open": true
|
|
458
|
+
},
|
|
459
|
+
{
|
|
460
|
+
"id": "qwen3-30b-a3b-instruct-jul-2025",
|
|
461
|
+
"name": "Qwen3-30B-A3B-Instruct (Jul 2025)",
|
|
462
|
+
"date": "2025-07-29",
|
|
463
|
+
"score": 129,
|
|
464
|
+
"out": 0.3,
|
|
465
|
+
"orId": "qwen/qwen3-30b-a3b-instruct-2507",
|
|
466
|
+
"hf": "Qwen/Qwen3-30B-A3B-Instruct-2507",
|
|
467
|
+
"open": true
|
|
468
|
+
},
|
|
469
|
+
{
|
|
470
|
+
"id": "claude-opus-4-1",
|
|
471
|
+
"name": "Claude Opus 4.1",
|
|
472
|
+
"date": "2025-08-05",
|
|
473
|
+
"score": 141.1,
|
|
474
|
+
"out": 75,
|
|
475
|
+
"orId": "anthropic/claude-opus-4.1",
|
|
476
|
+
"open": false
|
|
477
|
+
},
|
|
478
|
+
{
|
|
479
|
+
"id": "gpt-oss-120b",
|
|
480
|
+
"name": "gpt-oss-120b",
|
|
481
|
+
"date": "2025-08-05",
|
|
482
|
+
"score": 132,
|
|
483
|
+
"out": 0.6,
|
|
484
|
+
"orId": "openai/gpt-oss-120b",
|
|
485
|
+
"hf": "openai/gpt-oss-120b",
|
|
486
|
+
"open": true
|
|
487
|
+
},
|
|
488
|
+
{
|
|
489
|
+
"id": "gpt-oss-20b",
|
|
490
|
+
"name": "gpt-oss-20b",
|
|
491
|
+
"date": "2025-08-05",
|
|
492
|
+
"score": 129.6,
|
|
493
|
+
"out": 0.09,
|
|
494
|
+
"orId": "openai/gpt-oss-20b",
|
|
495
|
+
"hf": "openai/gpt-oss-20b",
|
|
496
|
+
"open": true
|
|
497
|
+
},
|
|
498
|
+
{
|
|
499
|
+
"id": "gpt-5",
|
|
500
|
+
"name": "GPT-5",
|
|
501
|
+
"date": "2025-08-07",
|
|
502
|
+
"score": 150.9,
|
|
503
|
+
"out": 10,
|
|
504
|
+
"orId": "openai/gpt-5",
|
|
505
|
+
"open": false
|
|
506
|
+
},
|
|
507
|
+
{
|
|
508
|
+
"id": "gpt-5-mini",
|
|
509
|
+
"name": "GPT-5 mini",
|
|
510
|
+
"date": "2025-08-07",
|
|
511
|
+
"score": 142.2,
|
|
512
|
+
"out": 2,
|
|
513
|
+
"orId": "openai/gpt-5-mini",
|
|
514
|
+
"open": false
|
|
515
|
+
},
|
|
516
|
+
{
|
|
517
|
+
"id": "gpt-5-nano",
|
|
518
|
+
"name": "GPT-5 nano",
|
|
519
|
+
"date": "2025-08-07",
|
|
520
|
+
"score": 127.2,
|
|
521
|
+
"out": 0.4,
|
|
522
|
+
"orId": "openai/gpt-5-nano",
|
|
523
|
+
"open": false
|
|
524
|
+
},
|
|
525
|
+
{
|
|
526
|
+
"id": "mistral-medium-3-1",
|
|
527
|
+
"name": "Mistral Medium 3.1",
|
|
528
|
+
"date": "2025-08-12",
|
|
529
|
+
"score": 120.2,
|
|
530
|
+
"out": 2,
|
|
531
|
+
"orId": "mistralai/mistral-medium-3.1",
|
|
532
|
+
"open": false
|
|
533
|
+
},
|
|
534
|
+
{
|
|
535
|
+
"id": "deepseek-v3-1",
|
|
536
|
+
"name": "DeepSeek-V3.1",
|
|
537
|
+
"date": "2025-08-21",
|
|
538
|
+
"score": 129.3,
|
|
539
|
+
"out": 0.95,
|
|
540
|
+
"orId": "deepseek/deepseek-chat-v3.1",
|
|
541
|
+
"hf": "deepseek-ai/DeepSeek-V3.1",
|
|
542
|
+
"open": true
|
|
543
|
+
},
|
|
544
|
+
{
|
|
545
|
+
"id": "kimi-k2-sep-2025",
|
|
546
|
+
"name": "Kimi K2 (Sep 2025)",
|
|
547
|
+
"date": "2025-09-05",
|
|
548
|
+
"score": 133.6,
|
|
549
|
+
"out": 2.5,
|
|
550
|
+
"orId": "moonshotai/kimi-k2-0905",
|
|
551
|
+
"hf": "moonshotai/Kimi-K2-Instruct-0905",
|
|
552
|
+
"open": true
|
|
553
|
+
},
|
|
554
|
+
{
|
|
555
|
+
"id": "deepseek-v3-1-terminus",
|
|
556
|
+
"name": "DeepSeek-V3.1-Terminus",
|
|
557
|
+
"date": "2025-09-22",
|
|
558
|
+
"score": 140.1,
|
|
559
|
+
"out": 1,
|
|
560
|
+
"orId": "deepseek/deepseek-v3.1-terminus",
|
|
561
|
+
"hf": "deepseek-ai/DeepSeek-V3.1-Terminus",
|
|
562
|
+
"open": true
|
|
563
|
+
},
|
|
564
|
+
{
|
|
565
|
+
"id": "claude-sonnet-4-5",
|
|
566
|
+
"name": "Claude Sonnet 4.5",
|
|
567
|
+
"date": "2025-09-29",
|
|
568
|
+
"score": 146.4,
|
|
569
|
+
"out": 15,
|
|
570
|
+
"orId": "anthropic/claude-sonnet-4.5",
|
|
571
|
+
"open": false
|
|
572
|
+
},
|
|
573
|
+
{
|
|
574
|
+
"id": "deepseek-v3-2-exp",
|
|
575
|
+
"name": "DeepSeek-V3.2-Exp",
|
|
576
|
+
"date": "2025-09-29",
|
|
577
|
+
"score": 142.4,
|
|
578
|
+
"out": 0.41,
|
|
579
|
+
"orId": "deepseek/deepseek-v3.2-exp",
|
|
580
|
+
"hf": "deepseek-ai/DeepSeek-V3.2-Exp",
|
|
581
|
+
"open": true
|
|
582
|
+
},
|
|
583
|
+
{
|
|
584
|
+
"id": "glm-4-6",
|
|
585
|
+
"name": "GLM-4.6",
|
|
586
|
+
"date": "2025-09-30",
|
|
587
|
+
"score": 130,
|
|
588
|
+
"out": 2,
|
|
589
|
+
"orId": "z-ai/glm-4.6",
|
|
590
|
+
"hf": "zai-org/GLM-4.6",
|
|
591
|
+
"open": true
|
|
592
|
+
},
|
|
593
|
+
{
|
|
594
|
+
"id": "claude-haiku-4-5",
|
|
595
|
+
"name": "Claude Haiku 4.5",
|
|
596
|
+
"date": "2025-10-15",
|
|
597
|
+
"score": 137.6,
|
|
598
|
+
"out": 5,
|
|
599
|
+
"orId": "anthropic/claude-haiku-4.5",
|
|
600
|
+
"open": false
|
|
601
|
+
},
|
|
602
|
+
{
|
|
603
|
+
"id": "minimax-m2",
|
|
604
|
+
"name": "MiniMax-M2",
|
|
605
|
+
"date": "2025-10-27",
|
|
606
|
+
"score": 139.8,
|
|
607
|
+
"out": 1.02,
|
|
608
|
+
"orId": "minimax/minimax-m2",
|
|
609
|
+
"hf": "MiniMaxAI/MiniMax-M2",
|
|
610
|
+
"open": true
|
|
611
|
+
},
|
|
612
|
+
{
|
|
613
|
+
"id": "kimi-k2-thinking",
|
|
614
|
+
"name": "Kimi K2 Thinking",
|
|
615
|
+
"date": "2025-11-06",
|
|
616
|
+
"score": 142.4,
|
|
617
|
+
"out": 2.5,
|
|
618
|
+
"orId": "moonshotai/kimi-k2-thinking",
|
|
619
|
+
"hf": "moonshotai/Kimi-K2-Thinking",
|
|
620
|
+
"open": true
|
|
621
|
+
},
|
|
622
|
+
{
|
|
623
|
+
"id": "gpt-5-1",
|
|
624
|
+
"name": "GPT-5.1",
|
|
625
|
+
"date": "2025-11-12",
|
|
626
|
+
"score": 152.3,
|
|
627
|
+
"out": 10,
|
|
628
|
+
"orId": "openai/gpt-5.1",
|
|
629
|
+
"open": false
|
|
630
|
+
},
|
|
631
|
+
{
|
|
632
|
+
"id": "gpt-5-1-codex",
|
|
633
|
+
"name": "GPT-5.1-Codex",
|
|
634
|
+
"date": "2025-11-12",
|
|
635
|
+
"score": 143.7,
|
|
636
|
+
"out": 10,
|
|
637
|
+
"orId": "openai/gpt-5.1-codex",
|
|
638
|
+
"open": false
|
|
639
|
+
},
|
|
640
|
+
{
|
|
641
|
+
"id": "gpt-5-1-codex-max",
|
|
642
|
+
"name": "GPT-5.1-Codex-Max",
|
|
643
|
+
"date": "2025-11-19",
|
|
644
|
+
"score": 146.5,
|
|
645
|
+
"out": 10,
|
|
646
|
+
"orId": "openai/gpt-5.1-codex-max",
|
|
647
|
+
"open": false
|
|
648
|
+
},
|
|
649
|
+
{
|
|
650
|
+
"id": "claude-opus-4-5",
|
|
651
|
+
"name": "Claude Opus 4.5",
|
|
652
|
+
"date": "2025-11-24",
|
|
653
|
+
"score": 153.5,
|
|
654
|
+
"out": 25,
|
|
655
|
+
"orId": "anthropic/claude-opus-4.5",
|
|
656
|
+
"open": false
|
|
657
|
+
},
|
|
658
|
+
{
|
|
659
|
+
"id": "deepseek-v3-2",
|
|
660
|
+
"name": "DeepSeek-V3.2",
|
|
661
|
+
"date": "2025-12-01",
|
|
662
|
+
"score": 143.7,
|
|
663
|
+
"out": 0.42,
|
|
664
|
+
"orId": "deepseek/deepseek-v3.2",
|
|
665
|
+
"hf": "deepseek-ai/DeepSeek-V3.2",
|
|
666
|
+
"open": true
|
|
667
|
+
},
|
|
668
|
+
{
|
|
669
|
+
"id": "mistral-large-3",
|
|
670
|
+
"name": "Mistral Large 3",
|
|
671
|
+
"date": "2025-12-02",
|
|
672
|
+
"score": 121.3,
|
|
673
|
+
"out": 1.5,
|
|
674
|
+
"orId": "mistralai/mistral-large-2512",
|
|
675
|
+
"open": true
|
|
676
|
+
},
|
|
677
|
+
{
|
|
678
|
+
"id": "devstral-2",
|
|
679
|
+
"name": "Devstral 2",
|
|
680
|
+
"date": "2025-12-09",
|
|
681
|
+
"score": 120.9,
|
|
682
|
+
"out": 2,
|
|
683
|
+
"orId": "mistralai/devstral-2512",
|
|
684
|
+
"hf": "mistralai/Devstral-2-123B-Instruct-2512",
|
|
685
|
+
"open": true
|
|
686
|
+
},
|
|
687
|
+
{
|
|
688
|
+
"id": "gpt-5-2",
|
|
689
|
+
"name": "GPT-5.2",
|
|
690
|
+
"date": "2025-12-11",
|
|
691
|
+
"score": 164.5,
|
|
692
|
+
"out": 14,
|
|
693
|
+
"orId": "openai/gpt-5.2",
|
|
694
|
+
"open": false
|
|
695
|
+
},
|
|
696
|
+
{
|
|
697
|
+
"id": "gemini-3-flash",
|
|
698
|
+
"name": "Gemini 3 Flash",
|
|
699
|
+
"date": "2025-12-17",
|
|
700
|
+
"score": 156.7,
|
|
701
|
+
"out": 3,
|
|
702
|
+
"orId": "google/gemini-3-flash-preview",
|
|
703
|
+
"open": false
|
|
704
|
+
},
|
|
705
|
+
{
|
|
706
|
+
"id": "gpt-5-2-codex",
|
|
707
|
+
"name": "GPT-5.2-Codex",
|
|
708
|
+
"date": "2025-12-18",
|
|
709
|
+
"score": 154.6,
|
|
710
|
+
"out": 14,
|
|
711
|
+
"orId": "openai/gpt-5.2-codex",
|
|
712
|
+
"open": false
|
|
713
|
+
},
|
|
714
|
+
{
|
|
715
|
+
"id": "glm-4-7",
|
|
716
|
+
"name": "GLM-4.7",
|
|
717
|
+
"date": "2025-12-22",
|
|
718
|
+
"score": 141.5,
|
|
719
|
+
"out": 1.98,
|
|
720
|
+
"orId": "z-ai/glm-4.7",
|
|
721
|
+
"hf": "zai-org/GLM-4.7",
|
|
722
|
+
"open": true
|
|
723
|
+
},
|
|
724
|
+
{
|
|
725
|
+
"id": "minimax-m2-1",
|
|
726
|
+
"name": "MiniMax-M2.1",
|
|
727
|
+
"date": "2025-12-23",
|
|
728
|
+
"score": 140.8,
|
|
729
|
+
"out": 1.2,
|
|
730
|
+
"orId": "minimax/minimax-m2.1",
|
|
731
|
+
"hf": "MiniMaxAI/MiniMax-M2.1",
|
|
732
|
+
"open": true
|
|
733
|
+
},
|
|
734
|
+
{
|
|
735
|
+
"id": "kimi-k2-5",
|
|
736
|
+
"name": "Kimi K2.5",
|
|
737
|
+
"date": "2026-01-27",
|
|
738
|
+
"score": 147.5,
|
|
739
|
+
"out": 2.5,
|
|
740
|
+
"orId": "moonshotai/kimi-k2.5",
|
|
741
|
+
"hf": "moonshotai/Kimi-K2.5",
|
|
742
|
+
"open": true
|
|
743
|
+
},
|
|
744
|
+
{
|
|
745
|
+
"id": "claude-opus-4-6",
|
|
746
|
+
"name": "Claude Opus 4.6",
|
|
747
|
+
"date": "2026-02-05",
|
|
748
|
+
"score": 167.4,
|
|
749
|
+
"out": 25,
|
|
750
|
+
"orId": "anthropic/claude-opus-4.6",
|
|
751
|
+
"open": false
|
|
752
|
+
},
|
|
753
|
+
{
|
|
754
|
+
"id": "gpt-5-3-codex",
|
|
755
|
+
"name": "GPT-5.3-Codex",
|
|
756
|
+
"date": "2026-02-05",
|
|
757
|
+
"score": 167.3,
|
|
758
|
+
"out": 14,
|
|
759
|
+
"orId": "openai/gpt-5.3-codex",
|
|
760
|
+
"open": false
|
|
761
|
+
},
|
|
762
|
+
{
|
|
763
|
+
"id": "glm-5",
|
|
764
|
+
"name": "GLM-5",
|
|
765
|
+
"date": "2026-02-11",
|
|
766
|
+
"score": 143.7,
|
|
767
|
+
"out": 3.2,
|
|
768
|
+
"orId": "z-ai/glm-5",
|
|
769
|
+
"hf": "zai-org/GLM-5",
|
|
770
|
+
"open": true
|
|
771
|
+
},
|
|
772
|
+
{
|
|
773
|
+
"id": "minimax-m2-5",
|
|
774
|
+
"name": "MiniMax-M2.5",
|
|
775
|
+
"date": "2026-02-12",
|
|
776
|
+
"score": 145.7,
|
|
777
|
+
"out": 1.2,
|
|
778
|
+
"orId": "minimax/minimax-m2.5",
|
|
779
|
+
"hf": "MiniMaxAI/MiniMax-M2.5",
|
|
780
|
+
"open": true
|
|
781
|
+
},
|
|
782
|
+
{
|
|
783
|
+
"id": "qwen3-5",
|
|
784
|
+
"name": "Qwen3.5",
|
|
785
|
+
"date": "2026-02-16",
|
|
786
|
+
"score": 145.9,
|
|
787
|
+
"out": 3.6,
|
|
788
|
+
"orId": "qwen/qwen3.5-397b-a17b",
|
|
789
|
+
"hf": "Qwen/Qwen3.5-397B-A17B",
|
|
790
|
+
"open": true
|
|
791
|
+
},
|
|
792
|
+
{
|
|
793
|
+
"id": "claude-sonnet-4-6",
|
|
794
|
+
"name": "Claude Sonnet 4.6",
|
|
795
|
+
"date": "2026-02-17",
|
|
796
|
+
"score": 161.2,
|
|
797
|
+
"out": 15,
|
|
798
|
+
"orId": "anthropic/claude-sonnet-4.6",
|
|
799
|
+
"open": false
|
|
800
|
+
},
|
|
801
|
+
{
|
|
802
|
+
"id": "grok-4-20",
|
|
803
|
+
"name": "Grok 4.20",
|
|
804
|
+
"date": "2026-02-17",
|
|
805
|
+
"score": 159.1,
|
|
806
|
+
"out": 2.5,
|
|
807
|
+
"orId": "x-ai/grok-4.20",
|
|
808
|
+
"open": false
|
|
809
|
+
},
|
|
810
|
+
{
|
|
811
|
+
"id": "gemini-3-1-pro",
|
|
812
|
+
"name": "Gemini 3.1 Pro",
|
|
813
|
+
"date": "2026-02-19",
|
|
814
|
+
"score": 164.7,
|
|
815
|
+
"out": 12,
|
|
816
|
+
"orId": "google/gemini-3.1-pro-preview",
|
|
817
|
+
"open": false
|
|
818
|
+
},
|
|
819
|
+
{
|
|
820
|
+
"id": "qwen3-5-122b-a10b",
|
|
821
|
+
"name": "Qwen3.5-122B-A10B",
|
|
822
|
+
"date": "2026-02-24",
|
|
823
|
+
"score": 145.6,
|
|
824
|
+
"out": 2.08,
|
|
825
|
+
"orId": "qwen/qwen3.5-122b-a10b",
|
|
826
|
+
"hf": "Qwen/Qwen3.5-122B-A10B",
|
|
827
|
+
"open": true
|
|
828
|
+
},
|
|
829
|
+
{
|
|
830
|
+
"id": "qwen3-5-27b",
|
|
831
|
+
"name": "Qwen3.5-27B",
|
|
832
|
+
"date": "2026-02-24",
|
|
833
|
+
"score": 139.2,
|
|
834
|
+
"out": 2.6,
|
|
835
|
+
"orId": "qwen/qwen3.5-27b",
|
|
836
|
+
"hf": "Qwen/Qwen3.5-27B",
|
|
837
|
+
"open": true
|
|
838
|
+
},
|
|
839
|
+
{
|
|
840
|
+
"id": "qwen3-5-35b-a3b",
|
|
841
|
+
"name": "Qwen3.5-35B-A3B",
|
|
842
|
+
"date": "2026-02-24",
|
|
843
|
+
"score": 136.5,
|
|
844
|
+
"out": 1,
|
|
845
|
+
"orId": "qwen/qwen3.5-35b-a3b",
|
|
846
|
+
"hf": "Qwen/Qwen3.5-35B-A3B",
|
|
847
|
+
"open": true
|
|
848
|
+
},
|
|
849
|
+
{
|
|
850
|
+
"id": "qwen3-5-9b",
|
|
851
|
+
"name": "Qwen3.5-9B",
|
|
852
|
+
"date": "2026-02-24",
|
|
853
|
+
"score": 132.1,
|
|
854
|
+
"out": 0.25,
|
|
855
|
+
"orId": "qwen/qwen3.5-9b",
|
|
856
|
+
"hf": "Qwen/Qwen3.5-9B",
|
|
857
|
+
"open": true
|
|
858
|
+
},
|
|
859
|
+
{
|
|
860
|
+
"id": "gemini-3-1-flash-lite",
|
|
861
|
+
"name": "Gemini 3.1 Flash-Lite",
|
|
862
|
+
"date": "2026-03-03",
|
|
863
|
+
"score": 139.7,
|
|
864
|
+
"out": 1.5,
|
|
865
|
+
"orId": "google/gemini-3.1-flash-lite",
|
|
866
|
+
"open": false
|
|
867
|
+
},
|
|
868
|
+
{
|
|
869
|
+
"id": "gpt-5-4",
|
|
870
|
+
"name": "GPT-5.4",
|
|
871
|
+
"date": "2026-03-05",
|
|
872
|
+
"score": 169.1,
|
|
873
|
+
"out": 15,
|
|
874
|
+
"orId": "openai/gpt-5.4",
|
|
875
|
+
"open": false
|
|
876
|
+
},
|
|
877
|
+
{
|
|
878
|
+
"id": "nemotron-3-super",
|
|
879
|
+
"name": "Nemotron 3 Super",
|
|
880
|
+
"date": "2026-03-11",
|
|
881
|
+
"score": 131.9,
|
|
882
|
+
"out": 2.4,
|
|
883
|
+
"orId": "nvidia/nemotron-3-super-120b-a12b",
|
|
884
|
+
"hf": "nvidia/NVIDIA-Nemotron-3-Super-120B-A12B-FP8",
|
|
885
|
+
"open": true
|
|
886
|
+
},
|
|
887
|
+
{
|
|
888
|
+
"id": "mistral-small-4",
|
|
889
|
+
"name": "Mistral Small 4",
|
|
890
|
+
"date": "2026-03-16",
|
|
891
|
+
"score": 120.2,
|
|
892
|
+
"out": 0.6,
|
|
893
|
+
"orId": "mistralai/mistral-small-2603",
|
|
894
|
+
"hf": "mistralai/Mistral-Small-4-119B-2603",
|
|
895
|
+
"open": true
|
|
896
|
+
},
|
|
897
|
+
{
|
|
898
|
+
"id": "gpt-5-4-mini",
|
|
899
|
+
"name": "GPT-5.4 Mini",
|
|
900
|
+
"date": "2026-03-17",
|
|
901
|
+
"score": 149.5,
|
|
902
|
+
"out": 4.5,
|
|
903
|
+
"orId": "openai/gpt-5.4-mini",
|
|
904
|
+
"open": false
|
|
905
|
+
},
|
|
906
|
+
{
|
|
907
|
+
"id": "gpt-5-4-nano",
|
|
908
|
+
"name": "GPT-5.4 Nano",
|
|
909
|
+
"date": "2026-03-17",
|
|
910
|
+
"score": 143.6,
|
|
911
|
+
"out": 1.25,
|
|
912
|
+
"orId": "openai/gpt-5.4-nano",
|
|
913
|
+
"open": false
|
|
914
|
+
},
|
|
915
|
+
{
|
|
916
|
+
"id": "minimax-m2-7",
|
|
917
|
+
"name": "MiniMax-M2.7",
|
|
918
|
+
"date": "2026-03-18",
|
|
919
|
+
"score": 135.2,
|
|
920
|
+
"out": 1.2,
|
|
921
|
+
"orId": "minimax/minimax-m2.7",
|
|
922
|
+
"hf": "MiniMaxAI/MiniMax-M2.7",
|
|
923
|
+
"open": true
|
|
924
|
+
},
|
|
925
|
+
{
|
|
926
|
+
"id": "gemma-4-26b-a4b",
|
|
927
|
+
"name": "Gemma 4 26B A4B",
|
|
928
|
+
"date": "2026-04-02",
|
|
929
|
+
"score": 137.1,
|
|
930
|
+
"out": 0.255,
|
|
931
|
+
"orId": "google/gemma-4-26b-a4b-it",
|
|
932
|
+
"hf": "google/gemma-4-26B-A4B-it",
|
|
933
|
+
"open": true
|
|
934
|
+
},
|
|
935
|
+
{
|
|
936
|
+
"id": "qwen3-6-plus",
|
|
937
|
+
"name": "Qwen3.6-Plus",
|
|
938
|
+
"date": "2026-04-02",
|
|
939
|
+
"score": 147.4,
|
|
940
|
+
"out": 1.95,
|
|
941
|
+
"orId": "qwen/qwen3.6-plus",
|
|
942
|
+
"open": false
|
|
943
|
+
},
|
|
944
|
+
{
|
|
945
|
+
"id": "glm-5-1",
|
|
946
|
+
"name": "GLM-5.1",
|
|
947
|
+
"date": "2026-04-07",
|
|
948
|
+
"score": 152.6,
|
|
949
|
+
"out": 4.4,
|
|
950
|
+
"orId": "z-ai/glm-5.1",
|
|
951
|
+
"hf": "zai-org/GLM-5.1",
|
|
952
|
+
"open": true
|
|
953
|
+
},
|
|
954
|
+
{
|
|
955
|
+
"id": "claude-opus-4-7",
|
|
956
|
+
"name": "Claude Opus 4.7",
|
|
957
|
+
"date": "2026-04-16",
|
|
958
|
+
"score": 169.4,
|
|
959
|
+
"out": 25,
|
|
960
|
+
"orId": "anthropic/claude-opus-4.7",
|
|
961
|
+
"open": false
|
|
962
|
+
},
|
|
963
|
+
{
|
|
964
|
+
"id": "qwen3-6-35b-a3b",
|
|
965
|
+
"name": "Qwen3.6-35B-A3B",
|
|
966
|
+
"date": "2026-04-16",
|
|
967
|
+
"score": 140.9,
|
|
968
|
+
"out": 0.9,
|
|
969
|
+
"orId": "qwen/qwen3.6-35b-a3b",
|
|
970
|
+
"hf": "Qwen/Qwen3.6-35B-A3B",
|
|
971
|
+
"open": true
|
|
972
|
+
},
|
|
973
|
+
{
|
|
974
|
+
"id": "grok-4-3",
|
|
975
|
+
"name": "Grok 4.3",
|
|
976
|
+
"date": "2026-04-17",
|
|
977
|
+
"score": 137.7,
|
|
978
|
+
"out": 2.5,
|
|
979
|
+
"orId": "x-ai/grok-4.3",
|
|
980
|
+
"open": false
|
|
981
|
+
},
|
|
982
|
+
{
|
|
983
|
+
"id": "kimi-k2-6",
|
|
984
|
+
"name": "Kimi K2.6",
|
|
985
|
+
"date": "2026-04-20",
|
|
986
|
+
"score": 155.2,
|
|
987
|
+
"out": 3.41,
|
|
988
|
+
"orId": "moonshotai/kimi-k2.6",
|
|
989
|
+
"hf": "moonshotai/Kimi-K2.6",
|
|
990
|
+
"open": true
|
|
991
|
+
},
|
|
992
|
+
{
|
|
993
|
+
"id": "qwen3-6-27b",
|
|
994
|
+
"name": "Qwen3.6 27B",
|
|
995
|
+
"date": "2026-04-22",
|
|
996
|
+
"score": 143.5,
|
|
997
|
+
"out": 3.2,
|
|
998
|
+
"orId": "qwen/qwen3.6-27b",
|
|
999
|
+
"hf": "Qwen/Qwen3.6-27B",
|
|
1000
|
+
"open": true
|
|
1001
|
+
},
|
|
1002
|
+
{
|
|
1003
|
+
"id": "gpt-5-5",
|
|
1004
|
+
"name": "GPT-5.5",
|
|
1005
|
+
"date": "2026-04-23",
|
|
1006
|
+
"score": 175.5,
|
|
1007
|
+
"out": 30,
|
|
1008
|
+
"orId": "openai/gpt-5.5",
|
|
1009
|
+
"open": false
|
|
1010
|
+
},
|
|
1011
|
+
{
|
|
1012
|
+
"id": "deepseek-v4-flash",
|
|
1013
|
+
"name": "DeepSeek-V4-Flash (Apr 2026)",
|
|
1014
|
+
"date": "2026-04-24",
|
|
1015
|
+
"score": 143.3,
|
|
1016
|
+
"out": 0.182,
|
|
1017
|
+
"orId": "deepseek/deepseek-v4-flash",
|
|
1018
|
+
"hf": "deepseek-ai/DeepSeek-V4-Flash",
|
|
1019
|
+
"open": true
|
|
1020
|
+
},
|
|
1021
|
+
{
|
|
1022
|
+
"id": "deepseek-v4-pro",
|
|
1023
|
+
"name": "DeepSeek-V4-Pro (Apr 2026)",
|
|
1024
|
+
"date": "2026-04-24",
|
|
1025
|
+
"score": 150.1,
|
|
1026
|
+
"out": 1.6325,
|
|
1027
|
+
"orId": "deepseek/deepseek-v4-pro",
|
|
1028
|
+
"hf": "deepseek-ai/DeepSeek-V4-Pro",
|
|
1029
|
+
"open": true
|
|
1030
|
+
},
|
|
1031
|
+
{
|
|
1032
|
+
"id": "mistral-medium-3-5",
|
|
1033
|
+
"name": "Mistral Medium 3.5",
|
|
1034
|
+
"date": "2026-04-28",
|
|
1035
|
+
"score": 131.7,
|
|
1036
|
+
"out": 7.5,
|
|
1037
|
+
"orId": "mistralai/mistral-medium-3-5",
|
|
1038
|
+
"open": true
|
|
1039
|
+
},
|
|
1040
|
+
{
|
|
1041
|
+
"id": "gemini-3-5-flash",
|
|
1042
|
+
"name": "Gemini 3.5 Flash",
|
|
1043
|
+
"date": "2026-05-19",
|
|
1044
|
+
"score": 164.3,
|
|
1045
|
+
"out": 9,
|
|
1046
|
+
"orId": "google/gemini-3.5-flash",
|
|
1047
|
+
"open": false
|
|
1048
|
+
},
|
|
1049
|
+
{
|
|
1050
|
+
"id": "qwen3-7-max",
|
|
1051
|
+
"name": "Qwen3.7-Max",
|
|
1052
|
+
"date": "2026-05-19",
|
|
1053
|
+
"score": 162.4,
|
|
1054
|
+
"out": 4.425,
|
|
1055
|
+
"orId": "qwen/qwen3.7-max",
|
|
1056
|
+
"open": false
|
|
1057
|
+
},
|
|
1058
|
+
{
|
|
1059
|
+
"id": "grok-build-0-1",
|
|
1060
|
+
"name": "Grok Build 0.1",
|
|
1061
|
+
"date": "2026-05-20",
|
|
1062
|
+
"score": 144,
|
|
1063
|
+
"out": 2,
|
|
1064
|
+
"orId": "x-ai/grok-build-0.1",
|
|
1065
|
+
"open": false
|
|
1066
|
+
},
|
|
1067
|
+
{
|
|
1068
|
+
"id": "claude-opus-4-8",
|
|
1069
|
+
"name": "Claude Opus 4.8",
|
|
1070
|
+
"date": "2026-05-28",
|
|
1071
|
+
"score": 173,
|
|
1072
|
+
"out": 25,
|
|
1073
|
+
"orId": "anthropic/claude-opus-4.8",
|
|
1074
|
+
"open": false
|
|
1075
|
+
},
|
|
1076
|
+
{
|
|
1077
|
+
"id": "minimax-m3",
|
|
1078
|
+
"name": "MiniMax-M3",
|
|
1079
|
+
"date": "2026-06-01",
|
|
1080
|
+
"score": 148.6,
|
|
1081
|
+
"out": 1.2,
|
|
1082
|
+
"orId": "minimax/minimax-m3",
|
|
1083
|
+
"hf": "MiniMaxAI/Minimax-M3",
|
|
1084
|
+
"open": true
|
|
1085
|
+
},
|
|
1086
|
+
{
|
|
1087
|
+
"id": "qwen3-7-plus",
|
|
1088
|
+
"name": "Qwen3.7-Plus",
|
|
1089
|
+
"date": "2026-06-01",
|
|
1090
|
+
"score": 150.6,
|
|
1091
|
+
"out": 1.28,
|
|
1092
|
+
"orId": "qwen/qwen3.7-plus",
|
|
1093
|
+
"open": false
|
|
1094
|
+
},
|
|
1095
|
+
{
|
|
1096
|
+
"id": "nemotron-3-ultra",
|
|
1097
|
+
"name": "Nemotron 3 Ultra",
|
|
1098
|
+
"date": "2026-06-04",
|
|
1099
|
+
"score": 143.2,
|
|
1100
|
+
"out": 2.4,
|
|
1101
|
+
"orId": "nvidia/nemotron-3-ultra-550b-a55b",
|
|
1102
|
+
"hf": "nvidia/NVIDIA-Nemotron-3-Ultra-550B-A55B-BF16",
|
|
1103
|
+
"open": true
|
|
1104
|
+
},
|
|
1105
|
+
{
|
|
1106
|
+
"id": "claude-fable-5",
|
|
1107
|
+
"name": "Claude Fable 5",
|
|
1108
|
+
"date": "2026-06-09",
|
|
1109
|
+
"score": 186.1,
|
|
1110
|
+
"out": 50,
|
|
1111
|
+
"orId": "anthropic/claude-fable-5",
|
|
1112
|
+
"open": false
|
|
1113
|
+
},
|
|
1114
|
+
{
|
|
1115
|
+
"id": "kimi-k2-7-code",
|
|
1116
|
+
"name": "Kimi K2.7 Code",
|
|
1117
|
+
"date": "2026-06-12",
|
|
1118
|
+
"score": 153.5,
|
|
1119
|
+
"out": 3.1,
|
|
1120
|
+
"orId": "moonshotai/kimi-k2.7-code",
|
|
1121
|
+
"hf": "moonshotai/Kimi-K2.7-Code",
|
|
1122
|
+
"open": true
|
|
1123
|
+
},
|
|
1124
|
+
{
|
|
1125
|
+
"id": "glm-5-2",
|
|
1126
|
+
"name": "GLM-5.2",
|
|
1127
|
+
"date": "2026-06-16",
|
|
1128
|
+
"score": 156.5,
|
|
1129
|
+
"out": 7,
|
|
1130
|
+
"orId": "z-ai/glm-5.2",
|
|
1131
|
+
"hf": "zai-org/GLM-5.2",
|
|
1132
|
+
"open": true
|
|
1133
|
+
},
|
|
1134
|
+
{
|
|
1135
|
+
"id": "claude-sonnet-5",
|
|
1136
|
+
"name": "Claude Sonnet 5",
|
|
1137
|
+
"date": "2026-06-30",
|
|
1138
|
+
"score": 166.8,
|
|
1139
|
+
"out": 10,
|
|
1140
|
+
"orId": "anthropic/claude-sonnet-5",
|
|
1141
|
+
"open": false
|
|
1142
|
+
},
|
|
1143
|
+
{
|
|
1144
|
+
"id": "grok-4-5",
|
|
1145
|
+
"name": "Grok 4.5",
|
|
1146
|
+
"date": "2026-07-08",
|
|
1147
|
+
"score": 163.4,
|
|
1148
|
+
"out": 6,
|
|
1149
|
+
"orId": "x-ai/grok-4.5",
|
|
1150
|
+
"open": false
|
|
1151
|
+
},
|
|
1152
|
+
{
|
|
1153
|
+
"id": "gpt-5-6-luna",
|
|
1154
|
+
"name": "GPT-5.6 Luna",
|
|
1155
|
+
"date": "2026-07-09",
|
|
1156
|
+
"score": 166.2,
|
|
1157
|
+
"out": 1.2,
|
|
1158
|
+
"orId": "openai/gpt-5.6-luna",
|
|
1159
|
+
"open": false
|
|
1160
|
+
},
|
|
1161
|
+
{
|
|
1162
|
+
"id": "gpt-5-6-sol",
|
|
1163
|
+
"name": "GPT-5.6 Sol",
|
|
1164
|
+
"date": "2026-07-09",
|
|
1165
|
+
"score": 182.8,
|
|
1166
|
+
"out": 20,
|
|
1167
|
+
"orId": "openai/gpt-5.6-sol",
|
|
1168
|
+
"open": false
|
|
1169
|
+
},
|
|
1170
|
+
{
|
|
1171
|
+
"id": "gpt-5-6-terra",
|
|
1172
|
+
"name": "GPT-5.6 Terra",
|
|
1173
|
+
"date": "2026-07-09",
|
|
1174
|
+
"score": 175.6,
|
|
1175
|
+
"out": 12,
|
|
1176
|
+
"orId": "openai/gpt-5.6-terra",
|
|
1177
|
+
"open": false
|
|
1178
|
+
},
|
|
1179
|
+
{
|
|
1180
|
+
"id": "muse-spark-1-1",
|
|
1181
|
+
"name": "Muse Spark 1.1",
|
|
1182
|
+
"date": "2026-07-09",
|
|
1183
|
+
"score": 164.3,
|
|
1184
|
+
"out": 4.25,
|
|
1185
|
+
"orId": "meta/muse-spark-1.1",
|
|
1186
|
+
"open": false
|
|
1187
|
+
},
|
|
1188
|
+
{
|
|
1189
|
+
"id": "inkling",
|
|
1190
|
+
"name": "Inkling",
|
|
1191
|
+
"date": "2026-07-15",
|
|
1192
|
+
"score": 151,
|
|
1193
|
+
"out": 4.05,
|
|
1194
|
+
"orId": "thinkingmachines/inkling",
|
|
1195
|
+
"hf": "thinkingmachines/Inkling",
|
|
1196
|
+
"open": true
|
|
1197
|
+
},
|
|
1198
|
+
{
|
|
1199
|
+
"id": "kimi-k3",
|
|
1200
|
+
"name": "Kimi K3",
|
|
1201
|
+
"date": "2026-07-16",
|
|
1202
|
+
"score": 170.4,
|
|
1203
|
+
"out": 15,
|
|
1204
|
+
"orId": "moonshotai/kimi-k3",
|
|
1205
|
+
"hf": "moonshotai/Kimi-K3",
|
|
1206
|
+
"open": true
|
|
1207
|
+
},
|
|
1208
|
+
{
|
|
1209
|
+
"id": "gemini-3-5-flash-lite",
|
|
1210
|
+
"name": "Gemini 3.5 Flash-Lite",
|
|
1211
|
+
"date": "2026-07-21",
|
|
1212
|
+
"score": 142.4,
|
|
1213
|
+
"out": 2.5,
|
|
1214
|
+
"orId": "google/gemini-3.5-flash-lite",
|
|
1215
|
+
"open": false
|
|
1216
|
+
},
|
|
1217
|
+
{
|
|
1218
|
+
"id": "gemini-3-6-flash",
|
|
1219
|
+
"name": "Gemini 3.6 Flash",
|
|
1220
|
+
"date": "2026-07-21",
|
|
1221
|
+
"score": 163.5,
|
|
1222
|
+
"out": 3.75,
|
|
1223
|
+
"orId": "google/gemini-3.6-flash",
|
|
1224
|
+
"open": false
|
|
1225
|
+
},
|
|
1226
|
+
{
|
|
1227
|
+
"id": "claude-opus-5",
|
|
1228
|
+
"name": "Claude Opus 5",
|
|
1229
|
+
"date": "2026-07-24",
|
|
1230
|
+
"score": 187.6,
|
|
1231
|
+
"out": 25,
|
|
1232
|
+
"orId": "anthropic/claude-opus-5",
|
|
1233
|
+
"open": false
|
|
1234
|
+
},
|
|
1235
|
+
{
|
|
1236
|
+
"id": "qwen3-7-flash",
|
|
1237
|
+
"name": "Qwen3.7 Flash",
|
|
1238
|
+
"date": "2026-07-27",
|
|
1239
|
+
"score": 140.4,
|
|
1240
|
+
"out": 0.13,
|
|
1241
|
+
"orId": "qwen/qwen3.7-flash",
|
|
1242
|
+
"open": false
|
|
1243
|
+
},
|
|
1244
|
+
{
|
|
1245
|
+
"id": "inkling-small",
|
|
1246
|
+
"name": "Inkling-Small",
|
|
1247
|
+
"date": "2026-07-30",
|
|
1248
|
+
"score": 154.7,
|
|
1249
|
+
"out": 1.2,
|
|
1250
|
+
"orId": "thinkingmachines/inkling-small",
|
|
1251
|
+
"hf": "thinkingmachines/Inkling-Small",
|
|
1252
|
+
"open": true
|
|
1253
|
+
},
|
|
1254
|
+
{
|
|
1255
|
+
"id": "muse-spark-1-2",
|
|
1256
|
+
"name": "Muse Spark 1.2",
|
|
1257
|
+
"date": "2026-08-05",
|
|
1258
|
+
"score": 165,
|
|
1259
|
+
"out": 4.25,
|
|
1260
|
+
"orId": "meta/muse-spark-1.2",
|
|
1261
|
+
"open": false
|
|
1262
|
+
},
|
|
1263
|
+
{
|
|
1264
|
+
"id": "grok-4-6",
|
|
1265
|
+
"name": "Grok 4.6",
|
|
1266
|
+
"date": "2026-08-12",
|
|
1267
|
+
"score": 169.8,
|
|
1268
|
+
"out": 6,
|
|
1269
|
+
"orId": "x-ai/grok-4.6",
|
|
1270
|
+
"open": false
|
|
1271
|
+
},
|
|
1272
|
+
{
|
|
1273
|
+
"id": "deepseek-v4-pro-2026-08",
|
|
1274
|
+
"name": "DeepSeek-V4-Pro",
|
|
1275
|
+
"date": "2026-08-13",
|
|
1276
|
+
"score": 165.9,
|
|
1277
|
+
"out": 3.96,
|
|
1278
|
+
"orId": "deepseek/deepseek-v4-pro-0813",
|
|
1279
|
+
"hf": "deepseek-ai/DeepSeek-V4-Pro-0813",
|
|
1280
|
+
"open": true
|
|
1281
|
+
},
|
|
1282
|
+
{
|
|
1283
|
+
"id": "gemini-3-7-flash",
|
|
1284
|
+
"name": "Gemini 3.7 Flash",
|
|
1285
|
+
"date": "2026-08-13",
|
|
1286
|
+
"score": 173.2,
|
|
1287
|
+
"out": 3.75,
|
|
1288
|
+
"orId": "google/gemini-3.7-flash",
|
|
1289
|
+
"open": false
|
|
1290
|
+
},
|
|
1291
|
+
{
|
|
1292
|
+
"id": "glm-5-3",
|
|
1293
|
+
"name": "GLM-5.3",
|
|
1294
|
+
"date": "2026-08-14",
|
|
1295
|
+
"score": 170.6,
|
|
1296
|
+
"out": 3.6,
|
|
1297
|
+
"orId": "z-ai/glm-5.3",
|
|
1298
|
+
"hf": "zai-org/GLM-5.3",
|
|
1299
|
+
"open": true
|
|
1300
|
+
},
|
|
1301
|
+
{
|
|
1302
|
+
"id": "qwen3-8-27b",
|
|
1303
|
+
"name": "Qwen3.8-27B",
|
|
1304
|
+
"date": "2026-08-14",
|
|
1305
|
+
"score": 158.4,
|
|
1306
|
+
"out": 1.25,
|
|
1307
|
+
"orId": "qwen/qwen3.8-27b",
|
|
1308
|
+
"hf": "Qwen/Qwen3.8-27B",
|
|
1309
|
+
"open": true
|
|
1310
|
+
},
|
|
1311
|
+
{
|
|
1312
|
+
"id": "glm-5-3-flash",
|
|
1313
|
+
"name": "GLM-5.3-Flash",
|
|
1314
|
+
"date": "2026-08-20",
|
|
1315
|
+
"score": 163,
|
|
1316
|
+
"out": 0.0701,
|
|
1317
|
+
"orId": "z-ai/glm-5.3-flash",
|
|
1318
|
+
"hf": "zai-org/GLM-5.3-Flash",
|
|
1319
|
+
"open": true
|
|
1320
|
+
},
|
|
1321
|
+
{
|
|
1322
|
+
"id": "deepseek-v4-flash-vision-exp",
|
|
1323
|
+
"name": "DeepSeek-V4-Flash-Vision-Exp",
|
|
1324
|
+
"date": "2026-08-21",
|
|
1325
|
+
"score": 168,
|
|
1326
|
+
"out": 0.6468,
|
|
1327
|
+
"orId": "deepseek/deepseek-v4-flash-vision-exp",
|
|
1328
|
+
"hf": "deepseek-ai/DeepSeek-V4-Flash-Vision-Exp",
|
|
1329
|
+
"open": true
|
|
1330
|
+
},
|
|
1331
|
+
{
|
|
1332
|
+
"id": "hy4-preview",
|
|
1333
|
+
"name": "Hy4 preview",
|
|
1334
|
+
"date": "2026-08-28",
|
|
1335
|
+
"score": 173,
|
|
1336
|
+
"out": 2.501,
|
|
1337
|
+
"orId": "tencent/hy4-preview",
|
|
1338
|
+
"hf": "tencent/Hy4-preview",
|
|
1339
|
+
"open": true
|
|
1340
|
+
},
|
|
1341
|
+
{
|
|
1342
|
+
"id": "claude-fable-5-1",
|
|
1343
|
+
"name": "Claude Fable 5.1",
|
|
1344
|
+
"date": "2026-09-01",
|
|
1345
|
+
"score": 191.8,
|
|
1346
|
+
"out": 50,
|
|
1347
|
+
"orId": "anthropic/claude-fable-5.1",
|
|
1348
|
+
"open": false
|
|
1349
|
+
},
|
|
1350
|
+
{
|
|
1351
|
+
"id": "qwen3-8-max-0902",
|
|
1352
|
+
"name": "Qwen3.8 Max (0902)",
|
|
1353
|
+
"date": "2026-09-01",
|
|
1354
|
+
"score": 163.9,
|
|
1355
|
+
"out": 6,
|
|
1356
|
+
"orId": "qwen/qwen3.8-max-0902",
|
|
1357
|
+
"open": false
|
|
1358
|
+
},
|
|
1359
|
+
{
|
|
1360
|
+
"id": "gemini-3-8-flash",
|
|
1361
|
+
"name": "Gemini 3.8 Flash",
|
|
1362
|
+
"date": "2026-09-02",
|
|
1363
|
+
"score": 174.5,
|
|
1364
|
+
"out": 4.125,
|
|
1365
|
+
"orId": "google/gemini-3.8-flash",
|
|
1366
|
+
"open": false
|
|
1367
|
+
},
|
|
1368
|
+
{
|
|
1369
|
+
"id": "muse-spark-1-3",
|
|
1370
|
+
"name": "Muse Spark 1.3",
|
|
1371
|
+
"date": "2026-09-02",
|
|
1372
|
+
"score": 170.4,
|
|
1373
|
+
"out": 4.25,
|
|
1374
|
+
"orId": "meta/muse-spark-1.3",
|
|
1375
|
+
"open": false
|
|
1376
|
+
},
|
|
1377
|
+
{
|
|
1378
|
+
"id": "gpt-6-astra",
|
|
1379
|
+
"name": "GPT-6 Astra",
|
|
1380
|
+
"date": "2026-09-04",
|
|
1381
|
+
"score": 197.3,
|
|
1382
|
+
"out": 50,
|
|
1383
|
+
"orId": "openai/gpt-6-astra",
|
|
1384
|
+
"open": false
|
|
1385
|
+
},
|
|
1386
|
+
{
|
|
1387
|
+
"id": "deepseek-v4-1-flash",
|
|
1388
|
+
"name": "DeepSeek-V4.1-Flash",
|
|
1389
|
+
"date": "2026-09-10",
|
|
1390
|
+
"score": 166.6,
|
|
1391
|
+
"out": 0.5,
|
|
1392
|
+
"orId": "deepseek/deepseek-v4.1-flash",
|
|
1393
|
+
"hf": "deepseek-ai/DeepSeek-V4.1-Flash",
|
|
1394
|
+
"open": true
|
|
1395
|
+
},
|
|
1396
|
+
{
|
|
1397
|
+
"id": "step-5-preview",
|
|
1398
|
+
"name": "Step 5 Preview",
|
|
1399
|
+
"date": "2026-09-18",
|
|
1400
|
+
"score": 165.6,
|
|
1401
|
+
"out": 2.7,
|
|
1402
|
+
"orId": "stepfun/step-5-preview",
|
|
1403
|
+
"hf": "stepfun-ai/Step-5-Preview",
|
|
1404
|
+
"open": false
|
|
1405
|
+
},
|
|
1406
|
+
{
|
|
1407
|
+
"id": "grok-4-7",
|
|
1408
|
+
"name": "Grok 4.7",
|
|
1409
|
+
"date": "2026-09-21",
|
|
1410
|
+
"score": 165.7,
|
|
1411
|
+
"out": 6,
|
|
1412
|
+
"orId": "x-ai/grok-4.7",
|
|
1413
|
+
"open": false
|
|
1414
|
+
},
|
|
1415
|
+
{
|
|
1416
|
+
"id": "mimo-v2-6-flash",
|
|
1417
|
+
"name": "MiMo-V2.6-Flash",
|
|
1418
|
+
"date": "2026-09-21",
|
|
1419
|
+
"score": 170.7,
|
|
1420
|
+
"out": 0.28,
|
|
1421
|
+
"orId": "xiaomi/mimo-v2.6-flash",
|
|
1422
|
+
"hf": "XiaomiMiMo/MiMo-V2.6-Flash-RL",
|
|
1423
|
+
"open": true
|
|
1424
|
+
},
|
|
1425
|
+
{
|
|
1426
|
+
"id": "claude-opus-5-5",
|
|
1427
|
+
"name": "Claude Opus 5.5",
|
|
1428
|
+
"date": "2026-09-22",
|
|
1429
|
+
"score": 198,
|
|
1430
|
+
"out": 20,
|
|
1431
|
+
"orId": "anthropic/claude-opus-5.5",
|
|
1432
|
+
"open": false
|
|
1433
|
+
},
|
|
1434
|
+
{
|
|
1435
|
+
"id": "gpt-6-luna",
|
|
1436
|
+
"name": "GPT-6 Luna",
|
|
1437
|
+
"date": "2026-09-22",
|
|
1438
|
+
"score": 166,
|
|
1439
|
+
"out": 0.5,
|
|
1440
|
+
"orId": "openai/gpt-6-luna",
|
|
1441
|
+
"open": false
|
|
1442
|
+
},
|
|
1443
|
+
{
|
|
1444
|
+
"id": "gpt-6-sol",
|
|
1445
|
+
"name": "GPT-6 Sol",
|
|
1446
|
+
"date": "2026-09-22",
|
|
1447
|
+
"score": 182.2,
|
|
1448
|
+
"out": 10,
|
|
1449
|
+
"orId": "openai/gpt-6-sol",
|
|
1450
|
+
"open": false
|
|
1451
|
+
},
|
|
1452
|
+
{
|
|
1453
|
+
"id": "claude-sonnet-5-5",
|
|
1454
|
+
"name": "Claude Sonnet 5.5",
|
|
1455
|
+
"date": "2026-09-28",
|
|
1456
|
+
"score": 189.4,
|
|
1457
|
+
"out": 10,
|
|
1458
|
+
"orId": "anthropic/claude-sonnet-5.5",
|
|
1459
|
+
"open": false
|
|
1460
|
+
},
|
|
1461
|
+
{
|
|
1462
|
+
"id": "gpt-6-1-sol",
|
|
1463
|
+
"name": "GPT-6.1 Sol",
|
|
1464
|
+
"date": "2026-09-29",
|
|
1465
|
+
"score": 195,
|
|
1466
|
+
"out": 10,
|
|
1467
|
+
"orId": "openai/gpt-6.1-sol",
|
|
1468
|
+
"open": false
|
|
1469
|
+
},
|
|
1470
|
+
{
|
|
1471
|
+
"id": "gemini-4-argon",
|
|
1472
|
+
"name": "Gemini 4 Argon",
|
|
1473
|
+
"date": "2026-09-30",
|
|
1474
|
+
"score": 194.6,
|
|
1475
|
+
"out": 10,
|
|
1476
|
+
"orId": "google/gemini-4-argon",
|
|
1477
|
+
"open": false
|
|
1478
|
+
},
|
|
1479
|
+
{
|
|
1480
|
+
"id": "mistral-large-4",
|
|
1481
|
+
"name": "Mistral Large 4",
|
|
1482
|
+
"date": "2026-10-06",
|
|
1483
|
+
"score": 155.6,
|
|
1484
|
+
"out": 2.1,
|
|
1485
|
+
"orId": "mistralai/mistral-large-4-0",
|
|
1486
|
+
"hf": "mistralai/Mistral-Large-4-1T-A52B",
|
|
1487
|
+
"open": true
|
|
1488
|
+
},
|
|
1489
|
+
{
|
|
1490
|
+
"id": "claude-haiku-5-5",
|
|
1491
|
+
"name": "Claude Haiku 5.5",
|
|
1492
|
+
"date": "2026-10-07",
|
|
1493
|
+
"score": 170.9,
|
|
1494
|
+
"out": 0.5,
|
|
1495
|
+
"orId": "anthropic/claude-haiku-5.5",
|
|
1496
|
+
"open": false
|
|
1497
|
+
},
|
|
1498
|
+
{
|
|
1499
|
+
"id": "muse-wonder",
|
|
1500
|
+
"name": "Muse Wonder",
|
|
1501
|
+
"date": "2026-10-12",
|
|
1502
|
+
"score": 180.5,
|
|
1503
|
+
"out": 20,
|
|
1504
|
+
"orId": "meta/muse-wonder",
|
|
1505
|
+
"open": false
|
|
1506
|
+
},
|
|
1507
|
+
{
|
|
1508
|
+
"id": "muse-spark-open",
|
|
1509
|
+
"name": "Muse Spark Open",
|
|
1510
|
+
"date": "2026-10-15",
|
|
1511
|
+
"score": 164.4,
|
|
1512
|
+
"out": 4.25,
|
|
1513
|
+
"orId": "meta/muse-spark-open",
|
|
1514
|
+
"hf": "meta-models/Muse-Spark-Open-816B-A42B",
|
|
1515
|
+
"open": true
|
|
1516
|
+
},
|
|
1517
|
+
{
|
|
1518
|
+
"id": "deepseek-v4-1-flash-1017",
|
|
1519
|
+
"name": "DeepSeek V4.1 Flash 1017",
|
|
1520
|
+
"date": "2026-10-17",
|
|
1521
|
+
"score": 174.9,
|
|
1522
|
+
"out": 1.2,
|
|
1523
|
+
"orId": "deepseek/deepseek-v4.1-flash-1017",
|
|
1524
|
+
"hf": "deepseek-ai/DeepSeek-V4.1-Flash-1017",
|
|
1525
|
+
"open": true
|
|
1526
|
+
},
|
|
1527
|
+
{
|
|
1528
|
+
"id": "gpt-6-1-terra",
|
|
1529
|
+
"name": "GPT-6.1 Terra",
|
|
1530
|
+
"date": "2026-10-26",
|
|
1531
|
+
"score": 191.5,
|
|
1532
|
+
"out": 5,
|
|
1533
|
+
"orId": "openai/gpt-6.1-terra",
|
|
1534
|
+
"open": false
|
|
1535
|
+
}
|
|
1536
|
+
]
|
|
1537
|
+
}
|