CrispStrobe commited on
Commit
e2e9c3d
·
1 Parent(s): 54c2fe8

fix: aggressively parse raw config.json to estimate parameters for models with minified HF metadata

Browse files
Files changed (2) hide show
  1. data/providers.json +646 -646
  2. scripts/fetch-providers.js +42 -57
data/providers.json CHANGED
@@ -1116,6 +1116,18 @@
1116
  ],
1117
  "hf_id": "meta-llama/Meta-Llama-3-8B-Instruct"
1118
  },
 
 
 
 
 
 
 
 
 
 
 
 
1119
  {
1120
  "name": "openai-responses/gpt-5-nano",
1121
  "type": "vision",
@@ -1130,15 +1142,15 @@
1130
  "hf_private": true
1131
  },
1132
  {
1133
- "name": "alibaba/qwen-turbo",
1134
  "type": "chat",
1135
  "input_price_per_1m": 0.05,
1136
- "output_price_per_1m": 0.2,
1137
  "currency": "USD",
1138
- "context_window": 1000000,
1139
  "capabilities": [
1140
- "tools"
1141
  ],
 
1142
  "hf_private": true
1143
  },
1144
  {
@@ -1172,18 +1184,6 @@
1172
  "context_window": 8192,
1173
  "size_b": 8
1174
  },
1175
- {
1176
- "name": "azure/gpt-5-nano",
1177
- "type": "chat",
1178
- "input_price_per_1m": 0.05,
1179
- "output_price_per_1m": 0.4,
1180
- "currency": "USD",
1181
- "capabilities": [
1182
- "reasoning"
1183
- ],
1184
- "context_window": 200000,
1185
- "hf_private": true
1186
- },
1187
  {
1188
  "name": "together/meta-llama/Llama-3.2-3B-Instruct-Turbo",
1189
  "type": "chat",
@@ -1259,32 +1259,6 @@
1259
  "context_window": 32768,
1260
  "hf_private": true
1261
  },
1262
- {
1263
- "name": "openai-responses/gpt-4.1-nano",
1264
- "type": "vision",
1265
- "input_price_per_1m": 0.1,
1266
- "output_price_per_1m": 0.4,
1267
- "currency": "USD",
1268
- "capabilities": [
1269
- "vision"
1270
- ],
1271
- "context_window": 1047576,
1272
- "hf_private": true
1273
- },
1274
- {
1275
- "name": "deepinfra/Qwen/Qwen3-32B",
1276
- "type": "chat",
1277
- "input_price_per_1m": 0.1,
1278
- "output_price_per_1m": 0.3,
1279
- "currency": "USD",
1280
- "context_window": 40000,
1281
- "size_b": 32,
1282
- "capabilities": [
1283
- "tools",
1284
- "reasoning"
1285
- ],
1286
- "hf_id": "Qwen/Qwen3-32B"
1287
- },
1288
  {
1289
  "name": "google/gemini-2.0-flash-001",
1290
  "type": "vision",
@@ -1311,20 +1285,16 @@
1311
  "hf_private": true
1312
  },
1313
  {
1314
- "name": "vertex/gemini-2.5-flash-lite",
1315
- "type": "vision",
1316
  "input_price_per_1m": 0.1,
1317
- "output_price_per_1m": 0.4,
1318
  "currency": "USD",
1319
- "capabilities": [
1320
- "vision",
1321
- "reasoning"
1322
- ],
1323
- "context_window": 1048576,
1324
- "hf_private": true
1325
  },
1326
  {
1327
- "name": "azure/gpt-4.1-nano",
1328
  "type": "vision",
1329
  "input_price_per_1m": 0.1,
1330
  "output_price_per_1m": 0.4,
@@ -1336,7 +1306,7 @@
1336
  "hf_private": true
1337
  },
1338
  {
1339
- "name": "azure/openai-responses/gpt-4.1-nano",
1340
  "type": "vision",
1341
  "input_price_per_1m": 0.1,
1342
  "output_price_per_1m": 0.4,
@@ -1348,13 +1318,18 @@
1348
  "hf_private": true
1349
  },
1350
  {
1351
- "name": "together/meta-llama/Meta-Llama-3-8B-Instruct-Lite",
1352
  "type": "chat",
1353
  "input_price_per_1m": 0.1,
1354
- "output_price_per_1m": 0.1,
1355
  "currency": "USD",
1356
- "context_window": 8192,
1357
- "size_b": 8
 
 
 
 
 
1358
  },
1359
  {
1360
  "name": "groq/openai/gpt-oss-20b",
@@ -1371,7 +1346,32 @@
1371
  "hf_id": "openai/gpt-oss-20b"
1372
  },
1373
  {
1374
- "name": "openai/gpt-4.1-nano",
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1375
  "type": "vision",
1376
  "input_price_per_1m": 0.1,
1377
  "output_price_per_1m": 0.4,
@@ -1458,15 +1458,6 @@
1458
  "context_window": 256000,
1459
  "hf_private": true
1460
  },
1461
- {
1462
- "name": "parasail/parasail-qwen3-235b-a22b-instruct-2507",
1463
- "type": "chat",
1464
- "input_price_per_1m": 0.15,
1465
- "output_price_per_1m": 0.85,
1466
- "currency": "USD",
1467
- "context_window": 262144,
1468
- "size_b": 235
1469
- },
1470
  {
1471
  "name": "nebius/openai/gpt-oss-120b",
1472
  "type": "chat",
@@ -1481,19 +1472,25 @@
1481
  "hf_id": "openai/gpt-oss-120b"
1482
  },
1483
  {
1484
- "name": "novita/deepseek/deepseek-r1-distill-qwen-14b",
1485
- "type": "chat",
1486
  "input_price_per_1m": 0.15,
1487
- "output_price_per_1m": 0.15,
1488
  "currency": "USD",
1489
- "context_window": 128000,
1490
- "size_b": 671,
1491
  "capabilities": [
1492
- "tools",
1493
- "reasoning"
1494
  ],
1495
- "hf_id": "deepseek-ai/DeepSeek-R1",
1496
- "size_source": "manual"
 
 
 
 
 
 
 
 
 
1497
  },
1498
  {
1499
  "name": "groq/openai/gpt-oss-120b",
@@ -1510,16 +1507,19 @@
1510
  "hf_id": "openai/gpt-oss-120b"
1511
  },
1512
  {
1513
- "name": "openai/gpt-4o-mini",
1514
- "type": "vision",
1515
  "input_price_per_1m": 0.15,
1516
- "output_price_per_1m": 0.6,
1517
  "currency": "USD",
 
 
1518
  "capabilities": [
1519
- "vision"
 
1520
  ],
1521
- "context_window": 128000,
1522
- "hf_private": true
1523
  },
1524
  {
1525
  "name": "novita/mistralai/mistral-nemo",
@@ -1544,86 +1544,81 @@
1544
  "size_b": 8
1545
  },
1546
  {
1547
- "name": "openai-responses/gpt-5.4-nano",
1548
  "type": "vision",
1549
  "input_price_per_1m": 0.2,
1550
- "output_price_per_1m": 1.25,
1551
  "currency": "USD",
1552
  "capabilities": [
1553
- "vision",
1554
- "reasoning"
1555
  ],
1556
- "context_window": 400000,
1557
- "hf_private": true
 
1558
  },
1559
  {
1560
- "name": "deepinfra/zai-org/GLM-4.5-Air",
1561
  "type": "chat",
1562
  "input_price_per_1m": 0.2,
1563
- "output_price_per_1m": 1.1,
1564
  "currency": "USD",
1565
- "context_window": 131072,
1566
  "capabilities": [
 
1567
  "tools",
1568
  "reasoning"
1569
  ],
1570
- "hf_id": "zai-org/GLM-4.5-Air",
1571
- "size_b": 110.5
1572
  },
1573
  {
1574
- "name": "deepinfra/Qwen/Qwen3-235B-A22B",
1575
  "type": "chat",
1576
  "input_price_per_1m": 0.2,
1577
- "output_price_per_1m": 0.6,
1578
  "currency": "USD",
1579
- "context_window": 40000,
1580
- "size_b": 235,
1581
  "capabilities": [
 
1582
  "tools",
1583
  "reasoning"
1584
  ],
1585
- "hf_id": "Qwen/Qwen3-235B-A22B"
1586
  },
1587
  {
1588
- "name": "alibaba/qwen3-30b-a3b-instruct-2507",
1589
  "type": "vision",
1590
  "input_price_per_1m": 0.2,
1591
- "output_price_per_1m": 0.8,
1592
  "currency": "USD",
1593
  "capabilities": [
1594
  "vision"
1595
  ],
1596
- "context_window": 131072,
1597
- "size_b": 30,
1598
- "hf_id": "Qwen/Qwen3-30B-A3B-Instruct-2507"
1599
  },
1600
  {
1601
- "name": "novita/qwen/qwen3-235b-a22b-fp8",
1602
- "type": "chat",
1603
  "input_price_per_1m": 0.2,
1604
- "output_price_per_1m": 0.8,
1605
  "currency": "USD",
1606
- "context_window": 128000,
1607
- "size_b": 235,
1608
  "capabilities": [
1609
- "tools",
1610
- "reasoning"
1611
  ],
1612
- "hf_id": "Qwen/Qwen3-235B-A22B"
 
1613
  },
1614
  {
1615
- "name": "novita/meta-llama/llama-4-maverick-17b-128e-instruct-fp8",
1616
- "type": "chat",
1617
  "input_price_per_1m": 0.2,
1618
- "output_price_per_1m": 0.85,
1619
  "currency": "USD",
1620
- "context_window": 1048576,
1621
- "size_b": 17,
1622
  "capabilities": [
1623
- "vision",
1624
- "tools"
1625
  ],
1626
- "hf_id": "meta-llama/Llama-4-Maverick-17B-128E-Instruct"
 
1627
  },
1628
  {
1629
  "name": "together/meta-llama/LlamaGuard-2-8b",
@@ -1635,7 +1630,7 @@
1635
  "size_b": 8
1636
  },
1637
  {
1638
- "name": "openai/gpt-5.4-nano",
1639
  "type": "vision",
1640
  "input_price_per_1m": 0.2,
1641
  "output_price_per_1m": 1.25,
@@ -1648,68 +1643,73 @@
1648
  "hf_private": true
1649
  },
1650
  {
1651
- "name": "xai/grok-4-fast-non-reasoning",
1652
- "type": "chat",
1653
  "input_price_per_1m": 0.2,
1654
- "output_price_per_1m": 0.5,
1655
  "currency": "USD",
1656
- "context_window": 2000000,
1657
  "capabilities": [
1658
  "vision",
1659
- "tools",
1660
  "reasoning"
1661
  ],
 
1662
  "hf_private": true
1663
  },
1664
  {
1665
- "name": "xai/grok-4-1-fast-non-reasoning",
1666
- "type": "vision",
1667
  "input_price_per_1m": 0.2,
1668
- "output_price_per_1m": 0.5,
1669
  "currency": "USD",
 
1670
  "capabilities": [
1671
- "vision"
 
1672
  ],
1673
- "context_window": 2000000,
1674
- "hf_private": true
1675
  },
1676
  {
1677
- "name": "xai/grok-4-fast",
1678
  "type": "chat",
1679
  "input_price_per_1m": 0.2,
1680
- "output_price_per_1m": 0.5,
1681
  "currency": "USD",
1682
- "context_window": 2000000,
 
1683
  "capabilities": [
1684
- "vision",
1685
  "tools",
1686
  "reasoning"
1687
  ],
1688
- "hf_private": true
1689
  },
1690
  {
1691
- "name": "xai/grok-code-fast-1",
1692
- "type": "vision",
1693
  "input_price_per_1m": 0.2,
1694
- "output_price_per_1m": 1.5,
1695
  "currency": "USD",
 
 
1696
  "capabilities": [
1697
- "vision"
 
1698
  ],
1699
- "context_window": 256000,
1700
- "hf_private": true
1701
  },
1702
  {
1703
- "name": "xai/grok-4-1-fast-reasoning",
1704
- "type": "vision",
1705
  "input_price_per_1m": 0.2,
1706
- "output_price_per_1m": 0.5,
1707
  "currency": "USD",
 
 
1708
  "capabilities": [
1709
- "vision"
 
1710
  ],
1711
- "context_window": 2000000,
1712
- "hf_private": true
1713
  },
1714
  {
1715
  "name": "deepinfra/Qwen/Qwen2.5-72B-Instruct",
@@ -1724,19 +1724,6 @@
1724
  ],
1725
  "hf_id": "Qwen/Qwen2.5-VL-72B-Instruct"
1726
  },
1727
- {
1728
- "name": "deepinfra/meta-llama/Llama-3.3-70B-Instruct",
1729
- "type": "chat",
1730
- "input_price_per_1m": 0.23,
1731
- "output_price_per_1m": 0.4,
1732
- "currency": "USD",
1733
- "context_window": 131072,
1734
- "size_b": 70,
1735
- "capabilities": [
1736
- "tools"
1737
- ],
1738
- "hf_id": "meta-llama/Llama-3.3-70B-Instruct"
1739
- },
1740
  {
1741
  "name": "deepinfra/meta-llama/Meta-Llama-3.1-70B-Instruct",
1742
  "type": "chat",
@@ -1759,6 +1746,19 @@
1759
  ],
1760
  "hf_id": "deepseek-ai/DeepSeek-R1-Distill-Llama-70B"
1761
  },
 
 
 
 
 
 
 
 
 
 
 
 
 
1762
  {
1763
  "name": "mistral/open-mistral-7b",
1764
  "type": "chat",
@@ -1769,28 +1769,28 @@
1769
  "size_b": 7
1770
  },
1771
  {
1772
- "name": "openai-responses/gpt-5-mini",
1773
  "type": "vision",
1774
  "input_price_per_1m": 0.25,
1775
- "output_price_per_1m": 2,
1776
  "currency": "USD",
1777
  "capabilities": [
1778
- "vision",
1779
- "reasoning"
1780
  ],
1781
- "context_window": 400000,
1782
  "hf_private": true
1783
  },
1784
  {
1785
- "name": "google/gemini-3.1-flash-lite-preview",
1786
  "type": "vision",
1787
  "input_price_per_1m": 0.25,
1788
- "output_price_per_1m": 1.5,
1789
  "currency": "USD",
1790
  "capabilities": [
1791
- "vision"
 
1792
  ],
1793
- "context_window": 1048576,
1794
  "hf_private": true
1795
  },
1796
  {
@@ -1831,37 +1831,36 @@
1831
  "size_b": 685.4
1832
  },
1833
  {
1834
- "name": "deepseek/deepseek-chat",
1835
  "type": "chat",
1836
  "input_price_per_1m": 0.28,
1837
  "output_price_per_1m": 0.42,
1838
  "currency": "USD",
1839
  "context_window": 128000,
1840
- "capabilities": [
1841
- "tools"
1842
- ],
1843
- "hf_id": "deepseek-ai/DeepSeek-V3",
1844
- "size_b": 671,
1845
- "size_source": "manual"
1846
  },
1847
  {
1848
- "name": "deepseek/deepseek-reasoner",
1849
  "type": "chat",
1850
  "input_price_per_1m": 0.28,
1851
  "output_price_per_1m": 0.42,
1852
  "currency": "USD",
1853
  "context_window": 128000,
1854
- "hf_private": true
 
 
 
 
 
1855
  },
1856
  {
1857
- "name": "coding/gemini-2.5-flash",
1858
  "type": "vision",
1859
  "input_price_per_1m": 0.3,
1860
- "output_price_per_1m": 2.5,
1861
  "currency": "USD",
1862
  "capabilities": [
1863
- "vision",
1864
- "reasoning"
1865
  ],
1866
  "context_window": 1048576,
1867
  "hf_private": true
@@ -1876,119 +1875,142 @@
1876
  "hf_private": true
1877
  },
1878
  {
1879
- "name": "deepinfra/deepseek-ai/DeepSeek-V3.1",
1880
  "type": "chat",
1881
  "input_price_per_1m": 0.3,
1882
- "output_price_per_1m": 1,
1883
  "currency": "USD",
1884
- "context_window": 163840,
1885
  "capabilities": [
1886
- "tools",
1887
  "reasoning"
1888
  ],
1889
- "hf_id": "deepseek-ai/DeepSeek-V3.1-Terminus",
1890
- "size_b": 671
 
1891
  },
1892
  {
1893
- "name": "parasail/parasail-gemma3-27b-it",
1894
  "type": "chat",
1895
  "input_price_per_1m": 0.3,
1896
- "output_price_per_1m": 0.5,
1897
  "currency": "USD",
1898
- "context_window": 128000,
1899
- "size_b": 27
 
 
 
 
1900
  },
1901
  {
1902
- "name": "alibaba/qwen3-coder-flash",
1903
  "type": "vision",
1904
  "input_price_per_1m": 0.3,
1905
- "output_price_per_1m": 1.5,
1906
  "currency": "USD",
1907
  "capabilities": [
1908
- "vision"
 
1909
  ],
1910
- "context_window": 1048576,
1911
- "hf_private": true
 
1912
  },
1913
  {
1914
- "name": "nebius/minimax/minimax-m2.1",
1915
- "type": "chat",
1916
  "input_price_per_1m": 0.3,
1917
  "output_price_per_1m": 1.2,
1918
  "currency": "USD",
1919
  "capabilities": [
 
1920
  "reasoning"
1921
  ],
1922
  "context_window": 200000,
1923
- "hf_id": "MiniMaxAI/MiniMax-M2.1",
1924
- "size_b": 228.7
1925
  },
1926
  {
1927
- "name": "nebius/deepseek-ai/DeepSeek-V3.2",
1928
  "type": "chat",
1929
  "input_price_per_1m": 0.3,
1930
- "output_price_per_1m": 0.45,
1931
  "currency": "USD",
 
1932
  "capabilities": [
 
1933
  "reasoning"
1934
  ],
1935
- "context_window": 164000,
1936
- "hf_id": "deepseek-ai/DeepSeek-V3.2",
1937
- "size_b": 685.4
1938
  },
1939
  {
1940
- "name": "novita/deepseek/deepseek-r1-distill-qwen-32b",
1941
  "type": "chat",
1942
  "input_price_per_1m": 0.3,
1943
- "output_price_per_1m": 0.3,
1944
  "currency": "USD",
1945
- "context_window": 12800,
1946
- "size_b": 32,
1947
  "capabilities": [
1948
  "reasoning"
1949
  ],
1950
- "hf_id": "deepseek-ai/DeepSeek-R1-Distill-Qwen-32B"
 
1951
  },
1952
  {
1953
- "name": "novita/MiniMax-M2.7",
1954
  "type": "vision",
1955
  "input_price_per_1m": 0.3,
1956
- "output_price_per_1m": 1.2,
1957
  "currency": "USD",
1958
  "capabilities": [
1959
  "vision",
1960
  "reasoning"
1961
  ],
1962
- "context_window": 200000,
1963
- "hf_id": "MiniMaxAI/MiniMax-M2.7",
1964
  "hf_private": true
1965
  },
1966
  {
1967
- "name": "novita/minimax-m2.7",
 
 
 
 
 
 
 
 
 
1968
  "type": "vision",
1969
  "input_price_per_1m": 0.3,
1970
- "output_price_per_1m": 1.2,
1971
  "currency": "USD",
1972
  "capabilities": [
1973
  "vision",
1974
  "reasoning"
1975
  ],
1976
- "context_window": 200000,
1977
- "hf_id": "MiniMaxAI/MiniMax-M2.7",
1978
  "hf_private": true
1979
  },
1980
  {
1981
- "name": "google/gemini-2.5-flash",
1982
- "type": "vision",
1983
  "input_price_per_1m": 0.3,
1984
- "output_price_per_1m": 2.5,
1985
  "currency": "USD",
 
1986
  "capabilities": [
1987
- "vision",
1988
  "reasoning"
1989
  ],
1990
- "context_window": 1048576,
1991
- "hf_private": true
 
 
 
 
 
 
 
 
 
1992
  },
1993
  {
1994
  "name": "vertex/gemini-2.5-flash",
@@ -2017,21 +2039,20 @@
2017
  "hf_private": true
2018
  },
2019
  {
2020
- "name": "minimaxi/MiniMax-M2",
2021
  "type": "chat",
2022
  "input_price_per_1m": 0.3,
2023
- "output_price_per_1m": 1.2,
2024
  "currency": "USD",
2025
- "context_window": 200000,
 
2026
  "capabilities": [
2027
- "tools",
2028
  "reasoning"
2029
  ],
2030
- "hf_id": "MiniMaxAI/MiniMax-M2",
2031
- "size_b": 228.7
2032
  },
2033
  {
2034
- "name": "minimaxi/MiniMax-M2.7",
2035
  "type": "vision",
2036
  "input_price_per_1m": 0.3,
2037
  "output_price_per_1m": 1.2,
@@ -2045,7 +2066,7 @@
2045
  "hf_private": true
2046
  },
2047
  {
2048
- "name": "minimaxi/MiniMax-M2.5",
2049
  "type": "vision",
2050
  "input_price_per_1m": 0.3,
2051
  "output_price_per_1m": 1.2,
@@ -2055,28 +2076,7 @@
2055
  "reasoning"
2056
  ],
2057
  "context_window": 200000,
2058
- "hf_id": "MiniMaxAI/MiniMax-M2.5",
2059
- "size_b": 228.7
2060
- },
2061
- {
2062
- "name": "together/Qwen/Qwen2.5-7B-Instruct-Turbo",
2063
- "type": "chat",
2064
- "input_price_per_1m": 0.3,
2065
- "output_price_per_1m": 0.3,
2066
- "currency": "USD",
2067
- "context_window": 32768,
2068
- "size_b": 7
2069
- },
2070
- {
2071
- "name": "xai/grok-3-mini",
2072
- "type": "chat",
2073
- "input_price_per_1m": 0.3,
2074
- "output_price_per_1m": 0.5,
2075
- "currency": "USD",
2076
- "capabilities": [
2077
- "reasoning"
2078
- ],
2079
- "context_window": 131072,
2080
  "hf_private": true
2081
  },
2082
  {
@@ -2104,15 +2104,27 @@
2104
  {
2105
  "name": "novita/meta-llama/llama-3.3-70b-instruct",
2106
  "type": "chat",
2107
- "input_price_per_1m": 0.39,
2108
- "output_price_per_1m": 0.39,
 
 
 
 
 
 
 
 
 
 
 
 
 
2109
  "currency": "USD",
2110
  "context_window": 131072,
2111
- "size_b": 70,
2112
  "capabilities": [
2113
  "tools"
2114
  ],
2115
- "hf_id": "meta-llama/Llama-3.3-70B-Instruct"
2116
  },
2117
  {
2118
  "name": "mistral/mistral-medium-latest",
@@ -2144,43 +2156,6 @@
2144
  ],
2145
  "hf_private": true
2146
  },
2147
- {
2148
- "name": "openai-responses/gpt-4.1-mini",
2149
- "type": "vision",
2150
- "input_price_per_1m": 0.4,
2151
- "output_price_per_1m": 1.6,
2152
- "currency": "USD",
2153
- "capabilities": [
2154
- "vision"
2155
- ],
2156
- "context_window": 1047576,
2157
- "hf_private": true
2158
- },
2159
- {
2160
- "name": "deepinfra/Qwen/Qwen3-Coder-480B-A35B-Instruct",
2161
- "type": "chat",
2162
- "input_price_per_1m": 0.4,
2163
- "output_price_per_1m": 1.6,
2164
- "currency": "USD",
2165
- "context_window": 262144,
2166
- "size_b": 480,
2167
- "capabilities": [
2168
- "tools"
2169
- ],
2170
- "hf_id": "Qwen/Qwen3-Coder-480B-A35B-Instruct"
2171
- },
2172
- {
2173
- "name": "alibaba/qwen-plus",
2174
- "type": "chat",
2175
- "input_price_per_1m": 0.4,
2176
- "output_price_per_1m": 1.2,
2177
- "currency": "USD",
2178
- "context_window": 131072,
2179
- "capabilities": [
2180
- "tools"
2181
- ],
2182
- "hf_private": true
2183
- },
2184
  {
2185
  "name": "nebius/zai-org/GLM-4.7",
2186
  "type": "chat",
@@ -2208,30 +2183,19 @@
2208
  "hf_id": "Qwen/Qwen3-Coder-480B-A35B-Instruct"
2209
  },
2210
  {
2211
- "name": "novita/deepseek/deepseek-v3-0324",
2212
- "type": "chat",
2213
  "input_price_per_1m": 0.4,
2214
- "output_price_per_1m": 1.3,
2215
  "currency": "USD",
2216
- "context_window": 128000,
2217
  "capabilities": [
2218
- "tools",
2219
- "reasoning"
2220
  ],
2221
- "size_b": 685,
2222
- "hf_id": "deepseek-ai/DeepSeek-V3-0324"
2223
- },
2224
- {
2225
- "name": "novita/deepseek/deepseek-v3-turbo",
2226
- "type": "chat",
2227
- "input_price_per_1m": 0.4,
2228
- "output_price_per_1m": 1.3,
2229
- "currency": "USD",
2230
- "context_window": 128000,
2231
  "hf_private": true
2232
  },
2233
  {
2234
- "name": "azure/gpt-4.1-mini",
2235
  "type": "vision",
2236
  "input_price_per_1m": 0.4,
2237
  "output_price_per_1m": 1.6,
@@ -2242,6 +2206,19 @@
2242
  "context_window": 1047576,
2243
  "hf_private": true
2244
  },
 
 
 
 
 
 
 
 
 
 
 
 
 
2245
  {
2246
  "name": "azure/openai-responses/gpt-4.1-mini",
2247
  "type": "vision",
@@ -2255,7 +2232,7 @@
2255
  "hf_private": true
2256
  },
2257
  {
2258
- "name": "openai/gpt-4.1-mini",
2259
  "type": "vision",
2260
  "input_price_per_1m": 0.4,
2261
  "output_price_per_1m": 1.6,
@@ -2267,29 +2244,39 @@
2267
  "hf_private": true
2268
  },
2269
  {
2270
- "name": "mistral/mistral-large-latest",
2271
  "type": "chat",
2272
- "input_price_per_1m": 0.5,
2273
- "output_price_per_1m": 1.5,
2274
  "currency": "USD",
2275
- "context_window": 131072,
 
 
 
 
 
 
 
 
 
2276
  "capabilities": [
2277
- "tools"
 
2278
  ],
2279
- "hf_private": true
 
2280
  },
2281
  {
2282
- "name": "nebius/moonshotai/Kimi-K2-Instruct",
2283
  "type": "chat",
2284
  "input_price_per_1m": 0.5,
2285
- "output_price_per_1m": 2.4,
2286
  "currency": "USD",
2287
- "context_window": 131000,
2288
  "capabilities": [
2289
  "tools"
2290
  ],
2291
- "hf_id": "moonshotai/Kimi-K2-Instruct",
2292
- "size_b": 1026.5
2293
  },
2294
  {
2295
  "name": "nebius/moonshotai/kimi-k2.5",
@@ -2305,6 +2292,19 @@
2305
  "hf_id": "moonshotai/Kimi-K2.5",
2306
  "size_b": 1058.6
2307
  },
 
 
 
 
 
 
 
 
 
 
 
 
 
2308
  {
2309
  "name": "nebius/deepseek-ai/DeepSeek-V3-0324",
2310
  "type": "chat",
@@ -2346,29 +2346,29 @@
2346
  "hf_private": true
2347
  },
2348
  {
2349
- "name": "vertex/gemini-3.1-flash-image-preview",
2350
  "type": "vision",
2351
  "input_price_per_1m": 0.5,
2352
- "output_price_per_1m": 2,
2353
  "currency": "USD",
2354
  "capabilities": [
2355
  "vision",
2356
  "reasoning"
2357
  ],
2358
- "context_window": 131072,
2359
  "hf_private": true
2360
  },
2361
  {
2362
- "name": "vertex/gemini-3-flash-preview",
2363
  "type": "vision",
2364
  "input_price_per_1m": 0.5,
2365
- "output_price_per_1m": 3,
2366
  "currency": "USD",
2367
  "capabilities": [
2368
  "vision",
2369
  "reasoning"
2370
  ],
2371
- "context_window": 1048576,
2372
  "hf_private": true
2373
  },
2374
  {
@@ -2423,20 +2423,6 @@
2423
  "hf_id": "moonshotai/Kimi-K2-Instruct",
2424
  "size_b": 1026.5
2425
  },
2426
- {
2427
- "name": "deepinfra/zai-org/GLM-4.5",
2428
- "type": "chat",
2429
- "input_price_per_1m": 0.6,
2430
- "output_price_per_1m": 2.2,
2431
- "currency": "USD",
2432
- "context_window": 131072,
2433
- "capabilities": [
2434
- "tools",
2435
- "reasoning"
2436
- ],
2437
- "hf_id": "zai-org/GLM-4.5",
2438
- "size_b": 358.3
2439
- },
2440
  {
2441
  "name": "zai/GLM-4.7",
2442
  "type": "chat",
@@ -2505,7 +2491,35 @@
2505
  "size_b": 358.3
2506
  },
2507
  {
2508
- "name": "novita/zai-org/glm-4.5",
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2509
  "type": "chat",
2510
  "input_price_per_1m": 0.6,
2511
  "output_price_per_1m": 2.2,
@@ -2533,18 +2547,18 @@
2533
  "size_b": 356.8
2534
  },
2535
  {
2536
- "name": "moonshot/kimi-k2.5",
2537
- "type": "vision",
2538
  "input_price_per_1m": 0.6,
2539
- "output_price_per_1m": 3,
2540
  "currency": "USD",
 
2541
  "capabilities": [
2542
- "vision",
2543
  "reasoning"
2544
  ],
2545
- "context_window": 262144,
2546
- "hf_id": "moonshotai/Kimi-K2.5",
2547
- "size_b": 1058.6
2548
  },
2549
  {
2550
  "name": "moonshot/kimi-k2-thinking-turbo",
@@ -2600,32 +2614,18 @@
2600
  "size_b": 1026.5
2601
  },
2602
  {
2603
- "name": "minimaxi/MiniMax-M2.7-highspeed",
2604
- "type": "vision",
2605
- "input_price_per_1m": 0.6,
2606
- "output_price_per_1m": 2.4,
2607
- "currency": "USD",
2608
- "capabilities": [
2609
- "vision",
2610
- "reasoning"
2611
- ],
2612
- "context_window": 200000,
2613
- "hf_id": "MiniMaxAI/MiniMax-M2.7",
2614
- "hf_private": true
2615
- },
2616
- {
2617
- "name": "minimaxi/MiniMax-M2.5-highspeed",
2618
  "type": "vision",
2619
  "input_price_per_1m": 0.6,
2620
- "output_price_per_1m": 2.4,
2621
  "currency": "USD",
2622
  "capabilities": [
2623
  "vision",
2624
  "reasoning"
2625
  ],
2626
- "context_window": 200000,
2627
- "hf_id": "MiniMaxAI/MiniMax-M2.5",
2628
- "size_b": 228.7
2629
  },
2630
  {
2631
  "name": "novita/microsoft/wizardlm-2-8x22b",
@@ -2712,15 +2712,6 @@
2712
  "context_window": 400000,
2713
  "hf_private": true
2714
  },
2715
- {
2716
- "name": "deepinfra/meta-llama/Meta-Llama-3.1-405B-Instruct",
2717
- "type": "chat",
2718
- "input_price_per_1m": 0.8,
2719
- "output_price_per_1m": 0.8,
2720
- "currency": "USD",
2721
- "context_window": 130815,
2722
- "size_b": 405
2723
- },
2724
  {
2725
  "name": "nebius/deepseek-ai/DeepSeek-R1-0528",
2726
  "type": "chat",
@@ -2735,6 +2726,15 @@
2735
  "hf_id": "deepseek-ai/DeepSeek-R1-0528",
2736
  "size_b": 684.5
2737
  },
 
 
 
 
 
 
 
 
 
2738
  {
2739
  "name": "novita/qwen/qwen2.5-vl-72b-instruct",
2740
  "type": "chat",
@@ -2758,22 +2758,8 @@
2758
  "size_b": 70,
2759
  "capabilities": [
2760
  "reasoning"
2761
- ],
2762
- "hf_id": "deepseek-ai/DeepSeek-R1-Distill-Llama-70B"
2763
- },
2764
- {
2765
- "name": "deepinfra/deepseek-ai/DeepSeek-V3",
2766
- "type": "chat",
2767
- "input_price_per_1m": 0.85,
2768
- "output_price_per_1m": 0.9,
2769
- "currency": "USD",
2770
- "context_window": 128000,
2771
- "capabilities": [
2772
- "tools",
2773
- "reasoning"
2774
- ],
2775
- "hf_id": "deepseek-ai/DeepSeek-V3.1-Terminus",
2776
- "size_b": 671
2777
  },
2778
  {
2779
  "name": "deepinfra/deepseek-ai/DeepSeek-R1",
@@ -2790,6 +2776,20 @@
2790
  "size_b": 671,
2791
  "size_source": "manual"
2792
  },
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2793
  {
2794
  "name": "alibaba/qwen3-max",
2795
  "type": "vision",
@@ -2803,26 +2803,26 @@
2803
  "hf_private": true
2804
  },
2805
  {
2806
- "name": "together/meta-llama/Meta-Llama-3.1-70B-Instruct-Turbo",
2807
  "type": "chat",
2808
  "input_price_per_1m": 0.88,
2809
  "output_price_per_1m": 0.88,
2810
  "currency": "USD",
2811
  "context_window": 131072,
2812
- "size_b": 70
 
 
 
 
2813
  },
2814
  {
2815
- "name": "together/meta-llama/Llama-3.3-70B-Instruct-Turbo",
2816
  "type": "chat",
2817
  "input_price_per_1m": 0.88,
2818
  "output_price_per_1m": 0.88,
2819
  "currency": "USD",
2820
  "context_window": 131072,
2821
- "size_b": 70,
2822
- "capabilities": [
2823
- "tools"
2824
- ],
2825
- "hf_id": "meta-llama/Llama-3.3-70B-Instruct"
2826
  },
2827
  {
2828
  "name": "novita/deepseek/deepseek_v3",
@@ -2847,18 +2847,6 @@
2847
  "context_window": 131072,
2848
  "hf_private": true
2849
  },
2850
- {
2851
- "name": "anthropic/claude-haiku-4-5",
2852
- "type": "vision",
2853
- "input_price_per_1m": 1,
2854
- "output_price_per_1m": 5,
2855
- "currency": "USD",
2856
- "capabilities": [
2857
- "vision"
2858
- ],
2859
- "context_window": 200000,
2860
- "hf_private": true
2861
- },
2862
  {
2863
  "name": "zai/GLM-5",
2864
  "type": "vision",
@@ -2886,61 +2874,61 @@
2886
  "hf_private": true
2887
  },
2888
  {
2889
- "name": "bedrock/claude-haiku-4-5",
2890
- "type": "vision",
2891
  "input_price_per_1m": 1,
2892
- "output_price_per_1m": 5,
2893
  "currency": "USD",
 
2894
  "capabilities": [
2895
  "vision"
2896
  ],
2897
- "context_window": 200000,
2898
  "hf_private": true
2899
  },
2900
  {
2901
- "name": "novita/GLM-5",
2902
- "type": "chat",
2903
  "input_price_per_1m": 1,
2904
- "output_price_per_1m": 3.2,
2905
  "currency": "USD",
2906
  "capabilities": [
2907
- "reasoning"
2908
  ],
2909
- "context_window": 202800,
2910
- "hf_id": "zai-org/GLM-5",
2911
- "size_b": 753.9
2912
  },
2913
  {
2914
- "name": "perplexity/sonar",
2915
- "type": "chat",
2916
  "input_price_per_1m": 1,
2917
- "output_price_per_1m": 1,
2918
  "currency": "USD",
2919
- "context_window": 131072,
2920
  "capabilities": [
2921
  "vision"
2922
  ],
 
2923
  "hf_private": true
2924
  },
2925
  {
2926
- "name": "vertex/claude-haiku-4-5",
2927
  "type": "vision",
2928
  "input_price_per_1m": 1,
2929
- "output_price_per_1m": 5,
2930
  "currency": "USD",
2931
  "capabilities": [
2932
- "vision"
 
2933
  ],
2934
  "context_window": 200000,
2935
  "hf_private": true
2936
  },
2937
  {
2938
- "name": "groq/moonshotai/Kimi-K2-Instruct-0905",
2939
  "type": "chat",
2940
  "input_price_per_1m": 1,
2941
  "output_price_per_1m": 3,
2942
  "currency": "USD",
2943
- "context_window": 256000,
2944
  "capabilities": [
2945
  "tools"
2946
  ],
@@ -2948,12 +2936,12 @@
2948
  "size_b": 1026.5
2949
  },
2950
  {
2951
- "name": "groq/moonshotai/kimi-k2-instruct",
2952
  "type": "chat",
2953
  "input_price_per_1m": 1,
2954
  "output_price_per_1m": 3,
2955
  "currency": "USD",
2956
- "context_window": 131072,
2957
  "capabilities": [
2958
  "tools"
2959
  ],
@@ -2961,18 +2949,30 @@
2961
  "size_b": 1026.5
2962
  },
2963
  {
2964
- "name": "openai/o3",
2965
  "type": "vision",
2966
  "input_price_per_1m": 1,
2967
- "output_price_per_1m": 4,
2968
  "currency": "USD",
2969
  "capabilities": [
2970
- "vision",
2971
- "reasoning"
2972
  ],
2973
  "context_window": 200000,
2974
  "hf_private": true
2975
  },
 
 
 
 
 
 
 
 
 
 
 
 
 
2976
  {
2977
  "name": "openai-responses/o3-mini",
2978
  "type": "chat",
@@ -2998,7 +2998,7 @@
2998
  "hf_private": true
2999
  },
3000
  {
3001
- "name": "azure/o4-mini",
3002
  "type": "chat",
3003
  "input_price_per_1m": 1.1,
3004
  "output_price_per_1m": 4.4,
@@ -3010,7 +3010,7 @@
3010
  "hf_private": true
3011
  },
3012
  {
3013
- "name": "openai/o3-mini",
3014
  "type": "chat",
3015
  "input_price_per_1m": 1.1,
3016
  "output_price_per_1m": 4.4,
@@ -3021,6 +3021,15 @@
3021
  "context_window": 200000,
3022
  "hf_private": true
3023
  },
 
 
 
 
 
 
 
 
 
3024
  {
3025
  "name": "moonshot/kimi-k2-turbo-preview",
3026
  "type": "chat",
@@ -3035,16 +3044,7 @@
3035
  "size_b": 1026.5
3036
  },
3037
  {
3038
- "name": "together/Qwen/Qwen2.5-72B-Instruct-Turbo",
3039
- "type": "chat",
3040
- "input_price_per_1m": 1.2,
3041
- "output_price_per_1m": 1.2,
3042
- "currency": "USD",
3043
- "context_window": 32768,
3044
- "size_b": 72
3045
- },
3046
- {
3047
- "name": "coding/gemini-2.5-pro",
3048
  "type": "vision",
3049
  "input_price_per_1m": 1.25,
3050
  "output_price_per_1m": 10,
@@ -3057,20 +3057,21 @@
3057
  "hf_private": true
3058
  },
3059
  {
3060
- "name": "openai-responses/gpt-5",
3061
- "type": "vision",
3062
  "input_price_per_1m": 1.25,
3063
- "output_price_per_1m": 10,
3064
  "currency": "USD",
 
3065
  "capabilities": [
3066
- "vision",
3067
  "reasoning"
3068
  ],
3069
- "context_window": 400000,
3070
- "hf_private": true
3071
  },
3072
  {
3073
- "name": "openai-responses/gpt-5.1",
3074
  "type": "vision",
3075
  "input_price_per_1m": 1.25,
3076
  "output_price_per_1m": 10,
@@ -3083,7 +3084,7 @@
3083
  "hf_private": true
3084
  },
3085
  {
3086
- "name": "openai-responses/gpt-5.1-codex",
3087
  "type": "vision",
3088
  "input_price_per_1m": 1.25,
3089
  "output_price_per_1m": 10,
@@ -3109,7 +3110,7 @@
3109
  "hf_private": true
3110
  },
3111
  {
3112
- "name": "google/gemini-2.5-pro",
3113
  "type": "vision",
3114
  "input_price_per_1m": 1.25,
3115
  "output_price_per_1m": 10,
@@ -3118,11 +3119,11 @@
3118
  "vision",
3119
  "reasoning"
3120
  ],
3121
- "context_window": 1048576,
3122
  "hf_private": true
3123
  },
3124
  {
3125
- "name": "vertex/gemini-2.5-pro",
3126
  "type": "vision",
3127
  "input_price_per_1m": 1.25,
3128
  "output_price_per_1m": 10,
@@ -3131,49 +3132,50 @@
3131
  "vision",
3132
  "reasoning"
3133
  ],
3134
- "context_window": 1048576,
3135
  "hf_private": true
3136
  },
3137
  {
3138
- "name": "azure/gpt-5",
3139
- "type": "chat",
3140
  "input_price_per_1m": 1.25,
3141
  "output_price_per_1m": 10,
3142
  "currency": "USD",
3143
  "capabilities": [
 
3144
  "reasoning"
3145
  ],
3146
- "context_window": 200000,
3147
  "hf_private": true
3148
  },
3149
  {
3150
- "name": "azure/gpt-5.1",
3151
- "type": "chat",
3152
  "input_price_per_1m": 1.25,
3153
  "output_price_per_1m": 10,
3154
  "currency": "USD",
3155
  "capabilities": [
 
3156
  "reasoning"
3157
  ],
3158
- "context_window": 200000,
3159
  "hf_private": true
3160
  },
3161
  {
3162
- "name": "together/deepseek-ai/DeepSeek-V3",
3163
- "type": "chat",
3164
  "input_price_per_1m": 1.25,
3165
- "output_price_per_1m": 1.25,
3166
  "currency": "USD",
3167
- "context_window": 131072,
3168
  "capabilities": [
3169
- "tools",
3170
  "reasoning"
3171
  ],
3172
- "hf_id": "deepseek-ai/DeepSeek-V3.1-Terminus",
3173
- "size_b": 671
3174
  },
3175
  {
3176
- "name": "openai/gpt-5-chat",
3177
  "type": "vision",
3178
  "input_price_per_1m": 1.25,
3179
  "output_price_per_1m": 10,
@@ -3182,37 +3184,35 @@
3182
  "vision",
3183
  "reasoning"
3184
  ],
3185
- "context_window": 128000,
3186
  "hf_private": true
3187
  },
3188
  {
3189
- "name": "openai/gpt-5.1",
3190
- "type": "vision",
3191
  "input_price_per_1m": 1.25,
3192
  "output_price_per_1m": 10,
3193
  "currency": "USD",
3194
  "capabilities": [
3195
- "vision",
3196
  "reasoning"
3197
  ],
3198
- "context_window": 400000,
3199
  "hf_private": true
3200
  },
3201
  {
3202
- "name": "openai/gpt-5.1-chat",
3203
- "type": "vision",
3204
  "input_price_per_1m": 1.25,
3205
  "output_price_per_1m": 10,
3206
  "currency": "USD",
3207
  "capabilities": [
3208
- "vision",
3209
  "reasoning"
3210
  ],
3211
- "context_window": 128000,
3212
  "hf_private": true
3213
  },
3214
  {
3215
- "name": "novita/sao10k/l31-70b-euryale-v2.2",
3216
  "type": "chat",
3217
  "input_price_per_1m": 1.48,
3218
  "output_price_per_1m": 1.48,
@@ -3221,7 +3221,7 @@
3221
  "size_b": 70
3222
  },
3223
  {
3224
- "name": "novita/sao10k/l3-70b-euryale-v2.1",
3225
  "type": "chat",
3226
  "input_price_per_1m": 1.48,
3227
  "output_price_per_1m": 1.48,
@@ -3242,7 +3242,7 @@
3242
  "hf_private": true
3243
  },
3244
  {
3245
- "name": "openai-responses/gpt-5.2",
3246
  "type": "vision",
3247
  "input_price_per_1m": 1.75,
3248
  "output_price_per_1m": 14,
@@ -3251,11 +3251,11 @@
3251
  "vision",
3252
  "reasoning"
3253
  ],
3254
- "context_window": 400000,
3255
  "hf_private": true
3256
  },
3257
  {
3258
- "name": "openai-responses/gpt-5.3-chat",
3259
  "type": "vision",
3260
  "input_price_per_1m": 1.75,
3261
  "output_price_per_1m": 14,
@@ -3264,7 +3264,7 @@
3264
  "vision",
3265
  "reasoning"
3266
  ],
3267
- "context_window": 128000,
3268
  "hf_private": true
3269
  },
3270
  {
@@ -3281,7 +3281,7 @@
3281
  "hf_private": true
3282
  },
3283
  {
3284
- "name": "openai-responses/gpt-5.3-codex",
3285
  "type": "vision",
3286
  "input_price_per_1m": 1.75,
3287
  "output_price_per_1m": 14,
@@ -3294,7 +3294,7 @@
3294
  "hf_private": true
3295
  },
3296
  {
3297
- "name": "azure/gpt-5.3-codex",
3298
  "type": "vision",
3299
  "input_price_per_1m": 1.75,
3300
  "output_price_per_1m": 14,
@@ -3307,19 +3307,20 @@
3307
  "hf_private": true
3308
  },
3309
  {
3310
- "name": "azure/gpt-5.2",
3311
- "type": "chat",
3312
  "input_price_per_1m": 1.75,
3313
  "output_price_per_1m": 14,
3314
  "currency": "USD",
3315
  "capabilities": [
 
3316
  "reasoning"
3317
  ],
3318
- "context_window": 200000,
3319
  "hf_private": true
3320
  },
3321
  {
3322
- "name": "azure/openai-responses/gpt-5.2-codex",
3323
  "type": "vision",
3324
  "input_price_per_1m": 1.75,
3325
  "output_price_per_1m": 14,
@@ -3328,24 +3329,23 @@
3328
  "vision",
3329
  "reasoning"
3330
  ],
3331
- "context_window": 400000,
3332
  "hf_private": true
3333
  },
3334
  {
3335
- "name": "openai/gpt-5.2",
3336
- "type": "vision",
3337
  "input_price_per_1m": 1.75,
3338
  "output_price_per_1m": 14,
3339
  "currency": "USD",
3340
  "capabilities": [
3341
- "vision",
3342
  "reasoning"
3343
  ],
3344
- "context_window": 400000,
3345
  "hf_private": true
3346
  },
3347
  {
3348
- "name": "openai/gpt-5.2-chat",
3349
  "type": "vision",
3350
  "input_price_per_1m": 1.75,
3351
  "output_price_per_1m": 14,
@@ -3354,11 +3354,11 @@
3354
  "vision",
3355
  "reasoning"
3356
  ],
3357
- "context_window": 128000,
3358
  "hf_private": true
3359
  },
3360
  {
3361
- "name": "openai/gpt-5.3-chat",
3362
  "type": "vision",
3363
  "input_price_per_1m": 1.75,
3364
  "output_price_per_1m": 14,
@@ -3367,28 +3367,29 @@
3367
  "vision",
3368
  "reasoning"
3369
  ],
3370
- "context_window": 128000,
3371
  "hf_private": true
3372
  },
3373
  {
3374
- "name": "mistral/pixtral-large-latest",
3375
  "type": "chat",
3376
  "input_price_per_1m": 2,
3377
- "output_price_per_1m": 5,
3378
  "currency": "USD",
3379
  "context_window": 131072,
 
 
 
 
3380
  "hf_private": true
3381
  },
3382
  {
3383
- "name": "openai-responses/gpt-4.1",
3384
- "type": "vision",
3385
  "input_price_per_1m": 2,
3386
- "output_price_per_1m": 8,
3387
  "currency": "USD",
3388
- "capabilities": [
3389
- "vision"
3390
- ],
3391
- "context_window": 1047576,
3392
  "hf_private": true
3393
  },
3394
  {
@@ -3401,33 +3402,29 @@
3401
  "hf_private": true
3402
  },
3403
  {
3404
- "name": "perplexity/sonar-reasoning-pro",
3405
  "type": "chat",
3406
  "input_price_per_1m": 2,
3407
- "output_price_per_1m": 8,
3408
  "currency": "USD",
3409
  "context_window": 131072,
3410
- "capabilities": [
3411
- "vision",
3412
- "reasoning"
3413
- ],
3414
  "hf_private": true
3415
  },
3416
  {
3417
- "name": "google/gemini-3.1-pro-preview",
3418
  "type": "vision",
3419
  "input_price_per_1m": 2,
3420
- "output_price_per_1m": 12,
3421
  "currency": "USD",
3422
  "capabilities": [
3423
  "vision",
3424
  "reasoning"
3425
  ],
3426
- "context_window": 1048576,
3427
  "hf_private": true
3428
  },
3429
  {
3430
- "name": "google/gemini-3-pro-preview",
3431
  "type": "vision",
3432
  "input_price_per_1m": 2,
3433
  "output_price_per_1m": 12,
@@ -3440,7 +3437,7 @@
3440
  "hf_private": true
3441
  },
3442
  {
3443
- "name": "vertex/gemini-3.1-pro-preview",
3444
  "type": "vision",
3445
  "input_price_per_1m": 2,
3446
  "output_price_per_1m": 12,
@@ -3453,33 +3450,32 @@
3453
  "hf_private": true
3454
  },
3455
  {
3456
- "name": "vertex/gemini-3-pro-preview",
3457
  "type": "vision",
3458
  "input_price_per_1m": 2,
3459
- "output_price_per_1m": 12,
3460
  "currency": "USD",
3461
  "capabilities": [
3462
- "vision",
3463
- "reasoning"
3464
  ],
3465
- "context_window": 1048576,
3466
  "hf_private": true
3467
  },
3468
  {
3469
- "name": "vertex/gemini-3-pro-image-preview",
3470
  "type": "vision",
3471
  "input_price_per_1m": 2,
3472
- "output_price_per_1m": 12,
3473
  "currency": "USD",
3474
  "capabilities": [
3475
  "vision",
3476
  "reasoning"
3477
  ],
3478
- "context_window": 1048576,
3479
  "hf_private": true
3480
  },
3481
  {
3482
- "name": "azure/openai-responses/gpt-4.1",
3483
  "type": "vision",
3484
  "input_price_per_1m": 2,
3485
  "output_price_per_1m": 8,
@@ -3491,62 +3487,66 @@
3491
  "hf_private": true
3492
  },
3493
  {
3494
- "name": "azure/gpt-4.1",
3495
  "type": "vision",
3496
  "input_price_per_1m": 2,
3497
- "output_price_per_1m": 8,
3498
  "currency": "USD",
3499
  "capabilities": [
3500
- "vision"
 
3501
  ],
3502
- "context_window": 1047576,
3503
  "hf_private": true
3504
  },
3505
  {
3506
- "name": "openai/gpt-4.1",
3507
  "type": "vision",
3508
  "input_price_per_1m": 2,
3509
- "output_price_per_1m": 8,
3510
  "currency": "USD",
3511
  "capabilities": [
3512
- "vision"
 
3513
  ],
3514
- "context_window": 1047576,
3515
  "hf_private": true
3516
  },
3517
  {
3518
- "name": "openai/o4-mini-deep-research",
3519
  "type": "vision",
3520
  "input_price_per_1m": 2,
3521
- "output_price_per_1m": 8,
3522
  "currency": "USD",
3523
  "capabilities": [
3524
  "vision",
3525
  "reasoning"
3526
  ],
3527
- "context_window": 100000,
3528
  "hf_private": true
3529
  },
3530
  {
3531
- "name": "xai/grok-4.2-beta",
3532
  "type": "vision",
3533
  "input_price_per_1m": 2,
3534
- "output_price_per_1m": 6,
3535
  "currency": "USD",
3536
  "capabilities": [
3537
- "vision",
3538
- "reasoning"
3539
  ],
3540
- "context_window": 2000000,
3541
  "hf_private": true
3542
  },
3543
  {
3544
- "name": "xai/grok-2-1212",
3545
- "type": "chat",
3546
  "input_price_per_1m": 2,
3547
- "output_price_per_1m": 10,
3548
  "currency": "USD",
3549
- "context_window": 131072,
 
 
 
3550
  "hf_private": true
3551
  },
3552
  {
@@ -3563,20 +3563,19 @@
3563
  "hf_private": true
3564
  },
3565
  {
3566
- "name": "azure/openai-responses/gpt-5.4",
3567
  "type": "vision",
3568
  "input_price_per_1m": 2.5,
3569
- "output_price_per_1m": 15,
3570
  "currency": "USD",
3571
  "capabilities": [
3572
- "vision",
3573
- "reasoning"
3574
  ],
3575
- "context_window": 1050000,
3576
  "hf_private": true
3577
  },
3578
  {
3579
- "name": "openai/gpt-4o",
3580
  "type": "vision",
3581
  "input_price_per_1m": 2.5,
3582
  "output_price_per_1m": 10,
@@ -3588,19 +3587,20 @@
3588
  "hf_private": true
3589
  },
3590
  {
3591
- "name": "openai/gpt-4o-2024-05-13",
3592
  "type": "vision",
3593
  "input_price_per_1m": 2.5,
3594
- "output_price_per_1m": 10,
3595
  "currency": "USD",
3596
  "capabilities": [
3597
- "vision"
 
3598
  ],
3599
- "context_window": 128000,
3600
  "hf_private": true
3601
  },
3602
  {
3603
- "name": "openai/gpt-4o-2024-08-06",
3604
  "type": "vision",
3605
  "input_price_per_1m": 2.5,
3606
  "output_price_per_1m": 10,
@@ -3624,7 +3624,7 @@
3624
  "hf_private": true
3625
  },
3626
  {
3627
- "name": "openai/gpt-5.4",
3628
  "type": "vision",
3629
  "input_price_per_1m": 2.5,
3630
  "output_price_per_1m": 15,
@@ -3637,20 +3637,19 @@
3637
  "hf_private": true
3638
  },
3639
  {
3640
- "name": "coding/claude-sonnet-4-20250514",
3641
- "type": "vision",
3642
  "input_price_per_1m": 3,
3643
  "output_price_per_1m": 15,
3644
  "currency": "USD",
 
3645
  "capabilities": [
3646
- "vision",
3647
- "reasoning"
3648
  ],
3649
- "context_window": 200000,
3650
  "hf_private": true
3651
  },
3652
  {
3653
- "name": "anthropic/claude-sonnet-4",
3654
  "type": "vision",
3655
  "input_price_per_1m": 3,
3656
  "output_price_per_1m": 15,
@@ -3663,7 +3662,7 @@
3663
  "hf_private": true
3664
  },
3665
  {
3666
- "name": "anthropic/claude-sonnet-4-6",
3667
  "type": "vision",
3668
  "input_price_per_1m": 3,
3669
  "output_price_per_1m": 15,
@@ -3676,7 +3675,7 @@
3676
  "hf_private": true
3677
  },
3678
  {
3679
- "name": "anthropic/claude-sonnet-4-5",
3680
  "type": "vision",
3681
  "input_price_per_1m": 3,
3682
  "output_price_per_1m": 15,
@@ -3689,16 +3688,15 @@
3689
  "hf_private": true
3690
  },
3691
  {
3692
- "name": "bedrock/claude-sonnet-4",
3693
  "type": "vision",
3694
  "input_price_per_1m": 3,
3695
  "output_price_per_1m": 15,
3696
  "currency": "USD",
3697
  "capabilities": [
3698
- "vision",
3699
- "reasoning"
3700
  ],
3701
- "context_window": 200000,
3702
  "hf_private": true
3703
  },
3704
  {
@@ -3715,7 +3713,7 @@
3715
  "hf_private": true
3716
  },
3717
  {
3718
- "name": "bedrock/claude-sonnet-4-6",
3719
  "type": "vision",
3720
  "input_price_per_1m": 3,
3721
  "output_price_per_1m": 15,
@@ -3724,7 +3722,7 @@
3724
  "vision",
3725
  "reasoning"
3726
  ],
3727
- "context_window": 1000000,
3728
  "hf_private": true
3729
  },
3730
  {
@@ -3741,32 +3739,35 @@
3741
  "hf_private": true
3742
  },
3743
  {
3744
- "name": "perplexity/sonar-pro",
3745
- "type": "chat",
3746
  "input_price_per_1m": 3,
3747
  "output_price_per_1m": 15,
3748
  "currency": "USD",
3749
- "context_window": 204800,
3750
  "capabilities": [
3751
- "vision"
 
3752
  ],
 
3753
  "hf_private": true
3754
  },
3755
  {
3756
- "name": "vertex/claude-3-7-sonnet",
3757
- "type": "vision",
3758
  "input_price_per_1m": 3,
3759
- "output_price_per_1m": 15,
3760
  "currency": "USD",
 
3761
  "capabilities": [
3762
- "vision",
3763
  "reasoning"
3764
  ],
3765
- "context_window": 200000,
3766
- "hf_private": true
 
3767
  },
3768
  {
3769
- "name": "vertex/claude-sonnet-4-5",
3770
  "type": "vision",
3771
  "input_price_per_1m": 3,
3772
  "output_price_per_1m": 15,
@@ -3779,19 +3780,20 @@
3779
  "hf_private": true
3780
  },
3781
  {
3782
- "name": "vertex/claude-3-5-sonnet",
3783
  "type": "vision",
3784
  "input_price_per_1m": 3,
3785
  "output_price_per_1m": 15,
3786
  "currency": "USD",
3787
  "capabilities": [
3788
- "vision"
 
3789
  ],
3790
  "context_window": 200000,
3791
  "hf_private": true
3792
  },
3793
  {
3794
- "name": "vertex/claude-sonnet-4",
3795
  "type": "vision",
3796
  "input_price_per_1m": 3,
3797
  "output_price_per_1m": 15,
@@ -3804,30 +3806,28 @@
3804
  "hf_private": true
3805
  },
3806
  {
3807
- "name": "together/deepseek-ai/DeepSeek-R1",
3808
- "type": "chat",
3809
  "input_price_per_1m": 3,
3810
- "output_price_per_1m": 7,
3811
  "currency": "USD",
3812
- "context_window": 64000,
3813
  "capabilities": [
3814
- "tools",
3815
- "reasoning"
3816
  ],
3817
- "hf_id": "deepseek-ai/DeepSeek-R1",
3818
- "size_b": 671,
3819
- "size_source": "manual"
3820
  },
3821
  {
3822
- "name": "xai/grok-4",
3823
  "type": "vision",
3824
  "input_price_per_1m": 3,
3825
  "output_price_per_1m": 15,
3826
  "currency": "USD",
3827
  "capabilities": [
3828
- "vision"
 
3829
  ],
3830
- "context_window": 256000,
3831
  "hf_private": true
3832
  },
3833
  {
@@ -3872,20 +3872,19 @@
3872
  "hf_private": true
3873
  },
3874
  {
3875
- "name": "bedrock/claude-opus-4-5",
3876
- "type": "vision",
3877
  "input_price_per_1m": 5,
3878
  "output_price_per_1m": 25,
3879
  "currency": "USD",
 
3880
  "capabilities": [
3881
- "vision",
3882
- "reasoning"
3883
  ],
3884
- "context_window": 200000,
3885
  "hf_private": true
3886
  },
3887
  {
3888
- "name": "bedrock/claude-opus-4-6",
3889
  "type": "vision",
3890
  "input_price_per_1m": 5,
3891
  "output_price_per_1m": 25,
@@ -3894,11 +3893,11 @@
3894
  "vision",
3895
  "reasoning"
3896
  ],
3897
- "context_window": 1000000,
3898
  "hf_private": true
3899
  },
3900
  {
3901
- "name": "vertex/claude-opus-4-5",
3902
  "type": "vision",
3903
  "input_price_per_1m": 5,
3904
  "output_price_per_1m": 25,
@@ -3907,44 +3906,45 @@
3907
  "vision",
3908
  "reasoning"
3909
  ],
3910
- "context_window": 200000,
3911
  "hf_private": true
3912
  },
3913
  {
3914
- "name": "vertex/claude-opus-4-6",
3915
  "type": "vision",
3916
  "input_price_per_1m": 5,
3917
- "output_price_per_1m": 25,
3918
  "currency": "USD",
3919
  "capabilities": [
3920
- "vision",
3921
- "reasoning"
3922
  ],
3923
- "context_window": 1000000,
3924
  "hf_private": true
3925
  },
3926
  {
3927
- "name": "openai/chatgpt-4o",
3928
  "type": "vision",
3929
  "input_price_per_1m": 5,
3930
- "output_price_per_1m": 15,
3931
  "currency": "USD",
3932
  "capabilities": [
3933
- "vision"
 
3934
  ],
3935
- "context_window": 128000,
3936
  "hf_private": true
3937
  },
3938
  {
3939
- "name": "xai/grok-3",
3940
- "type": "chat",
3941
  "input_price_per_1m": 5,
3942
  "output_price_per_1m": 25,
3943
  "currency": "USD",
3944
- "context_window": 131072,
3945
  "capabilities": [
3946
- "tools"
 
3947
  ],
 
3948
  "hf_private": true
3949
  },
3950
  {
@@ -3961,7 +3961,7 @@
3961
  "hf_private": true
3962
  },
3963
  {
3964
- "name": "coding/claude-opus-4-20250514",
3965
  "type": "vision",
3966
  "input_price_per_1m": 15,
3967
  "output_price_per_1m": 75,
@@ -3974,7 +3974,7 @@
3974
  "hf_private": true
3975
  },
3976
  {
3977
- "name": "anthropic/claude-opus-4-1",
3978
  "type": "vision",
3979
  "input_price_per_1m": 15,
3980
  "output_price_per_1m": 75,
@@ -3987,7 +3987,7 @@
3987
  "hf_private": true
3988
  },
3989
  {
3990
- "name": "anthropic/claude-opus-4",
3991
  "type": "vision",
3992
  "input_price_per_1m": 15,
3993
  "output_price_per_1m": 75,
@@ -4013,10 +4013,10 @@
4013
  "hf_private": true
4014
  },
4015
  {
4016
- "name": "bedrock/claude-opus-4",
4017
  "type": "vision",
4018
  "input_price_per_1m": 15,
4019
- "output_price_per_1m": 75,
4020
  "currency": "USD",
4021
  "capabilities": [
4022
  "vision",
@@ -4026,7 +4026,7 @@
4026
  "hf_private": true
4027
  },
4028
  {
4029
- "name": "vertex/claude-opus-4",
4030
  "type": "vision",
4031
  "input_price_per_1m": 15,
4032
  "output_price_per_1m": 75,
@@ -4039,7 +4039,7 @@
4039
  "hf_private": true
4040
  },
4041
  {
4042
- "name": "vertex/claude-opus-4-1",
4043
  "type": "vision",
4044
  "input_price_per_1m": 15,
4045
  "output_price_per_1m": 75,
@@ -4052,10 +4052,10 @@
4052
  "hf_private": true
4053
  },
4054
  {
4055
- "name": "openai/o1",
4056
  "type": "vision",
4057
  "input_price_per_1m": 15,
4058
- "output_price_per_1m": 60,
4059
  "currency": "USD",
4060
  "capabilities": [
4061
  "vision",
@@ -4091,7 +4091,7 @@
4091
  "hf_private": true
4092
  },
4093
  {
4094
- "name": "azure/openai-responses/gpt-5.4-pro",
4095
  "type": "vision",
4096
  "input_price_per_1m": 30,
4097
  "output_price_per_1m": 180,
@@ -4104,7 +4104,7 @@
4104
  "hf_private": true
4105
  },
4106
  {
4107
- "name": "openai/gpt-5.4-pro",
4108
  "type": "vision",
4109
  "input_price_per_1m": 30,
4110
  "output_price_per_1m": 180,
 
1116
  ],
1117
  "hf_id": "meta-llama/Meta-Llama-3-8B-Instruct"
1118
  },
1119
+ {
1120
+ "name": "alibaba/qwen-turbo",
1121
+ "type": "chat",
1122
+ "input_price_per_1m": 0.05,
1123
+ "output_price_per_1m": 0.2,
1124
+ "currency": "USD",
1125
+ "context_window": 1000000,
1126
+ "capabilities": [
1127
+ "tools"
1128
+ ],
1129
+ "hf_private": true
1130
+ },
1131
  {
1132
  "name": "openai-responses/gpt-5-nano",
1133
  "type": "vision",
 
1142
  "hf_private": true
1143
  },
1144
  {
1145
+ "name": "azure/gpt-5-nano",
1146
  "type": "chat",
1147
  "input_price_per_1m": 0.05,
1148
+ "output_price_per_1m": 0.4,
1149
  "currency": "USD",
 
1150
  "capabilities": [
1151
+ "reasoning"
1152
  ],
1153
+ "context_window": 200000,
1154
  "hf_private": true
1155
  },
1156
  {
 
1184
  "context_window": 8192,
1185
  "size_b": 8
1186
  },
 
 
 
 
 
 
 
 
 
 
 
 
1187
  {
1188
  "name": "together/meta-llama/Llama-3.2-3B-Instruct-Turbo",
1189
  "type": "chat",
 
1259
  "context_window": 32768,
1260
  "hf_private": true
1261
  },
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1262
  {
1263
  "name": "google/gemini-2.0-flash-001",
1264
  "type": "vision",
 
1285
  "hf_private": true
1286
  },
1287
  {
1288
+ "name": "together/meta-llama/Meta-Llama-3-8B-Instruct-Lite",
1289
+ "type": "chat",
1290
  "input_price_per_1m": 0.1,
1291
+ "output_price_per_1m": 0.1,
1292
  "currency": "USD",
1293
+ "context_window": 8192,
1294
+ "size_b": 8
 
 
 
 
1295
  },
1296
  {
1297
+ "name": "openai-responses/gpt-4.1-nano",
1298
  "type": "vision",
1299
  "input_price_per_1m": 0.1,
1300
  "output_price_per_1m": 0.4,
 
1306
  "hf_private": true
1307
  },
1308
  {
1309
+ "name": "openai/gpt-4.1-nano",
1310
  "type": "vision",
1311
  "input_price_per_1m": 0.1,
1312
  "output_price_per_1m": 0.4,
 
1318
  "hf_private": true
1319
  },
1320
  {
1321
+ "name": "deepinfra/Qwen/Qwen3-32B",
1322
  "type": "chat",
1323
  "input_price_per_1m": 0.1,
1324
+ "output_price_per_1m": 0.3,
1325
  "currency": "USD",
1326
+ "context_window": 40000,
1327
+ "size_b": 32,
1328
+ "capabilities": [
1329
+ "tools",
1330
+ "reasoning"
1331
+ ],
1332
+ "hf_id": "Qwen/Qwen3-32B"
1333
  },
1334
  {
1335
  "name": "groq/openai/gpt-oss-20b",
 
1346
  "hf_id": "openai/gpt-oss-20b"
1347
  },
1348
  {
1349
+ "name": "vertex/gemini-2.5-flash-lite",
1350
+ "type": "vision",
1351
+ "input_price_per_1m": 0.1,
1352
+ "output_price_per_1m": 0.4,
1353
+ "currency": "USD",
1354
+ "capabilities": [
1355
+ "vision",
1356
+ "reasoning"
1357
+ ],
1358
+ "context_window": 1048576,
1359
+ "hf_private": true
1360
+ },
1361
+ {
1362
+ "name": "azure/openai-responses/gpt-4.1-nano",
1363
+ "type": "vision",
1364
+ "input_price_per_1m": 0.1,
1365
+ "output_price_per_1m": 0.4,
1366
+ "currency": "USD",
1367
+ "capabilities": [
1368
+ "vision"
1369
+ ],
1370
+ "context_window": 1047576,
1371
+ "hf_private": true
1372
+ },
1373
+ {
1374
+ "name": "azure/gpt-4.1-nano",
1375
  "type": "vision",
1376
  "input_price_per_1m": 0.1,
1377
  "output_price_per_1m": 0.4,
 
1458
  "context_window": 256000,
1459
  "hf_private": true
1460
  },
 
 
 
 
 
 
 
 
 
1461
  {
1462
  "name": "nebius/openai/gpt-oss-120b",
1463
  "type": "chat",
 
1472
  "hf_id": "openai/gpt-oss-120b"
1473
  },
1474
  {
1475
+ "name": "openai/gpt-4o-mini",
1476
+ "type": "vision",
1477
  "input_price_per_1m": 0.15,
1478
+ "output_price_per_1m": 0.6,
1479
  "currency": "USD",
 
 
1480
  "capabilities": [
1481
+ "vision"
 
1482
  ],
1483
+ "context_window": 128000,
1484
+ "hf_private": true
1485
+ },
1486
+ {
1487
+ "name": "parasail/parasail-qwen3-235b-a22b-instruct-2507",
1488
+ "type": "chat",
1489
+ "input_price_per_1m": 0.15,
1490
+ "output_price_per_1m": 0.85,
1491
+ "currency": "USD",
1492
+ "context_window": 262144,
1493
+ "size_b": 235
1494
  },
1495
  {
1496
  "name": "groq/openai/gpt-oss-120b",
 
1507
  "hf_id": "openai/gpt-oss-120b"
1508
  },
1509
  {
1510
+ "name": "novita/deepseek/deepseek-r1-distill-qwen-14b",
1511
+ "type": "chat",
1512
  "input_price_per_1m": 0.15,
1513
+ "output_price_per_1m": 0.15,
1514
  "currency": "USD",
1515
+ "context_window": 128000,
1516
+ "size_b": 14,
1517
  "capabilities": [
1518
+ "tools",
1519
+ "reasoning"
1520
  ],
1521
+ "hf_id": "deepseek-ai/DeepSeek-R1",
1522
+ "size_source": "manual"
1523
  },
1524
  {
1525
  "name": "novita/mistralai/mistral-nemo",
 
1544
  "size_b": 8
1545
  },
1546
  {
1547
+ "name": "alibaba/qwen3-30b-a3b-instruct-2507",
1548
  "type": "vision",
1549
  "input_price_per_1m": 0.2,
1550
+ "output_price_per_1m": 0.8,
1551
  "currency": "USD",
1552
  "capabilities": [
1553
+ "vision"
 
1554
  ],
1555
+ "context_window": 131072,
1556
+ "size_b": 30,
1557
+ "hf_id": "Qwen/Qwen3-30B-A3B-Instruct-2507"
1558
  },
1559
  {
1560
+ "name": "xai/grok-4-fast",
1561
  "type": "chat",
1562
  "input_price_per_1m": 0.2,
1563
+ "output_price_per_1m": 0.5,
1564
  "currency": "USD",
1565
+ "context_window": 2000000,
1566
  "capabilities": [
1567
+ "vision",
1568
  "tools",
1569
  "reasoning"
1570
  ],
1571
+ "hf_private": true
 
1572
  },
1573
  {
1574
+ "name": "xai/grok-4-fast-non-reasoning",
1575
  "type": "chat",
1576
  "input_price_per_1m": 0.2,
1577
+ "output_price_per_1m": 0.5,
1578
  "currency": "USD",
1579
+ "context_window": 2000000,
 
1580
  "capabilities": [
1581
+ "vision",
1582
  "tools",
1583
  "reasoning"
1584
  ],
1585
+ "hf_private": true
1586
  },
1587
  {
1588
+ "name": "xai/grok-4-1-fast-reasoning",
1589
  "type": "vision",
1590
  "input_price_per_1m": 0.2,
1591
+ "output_price_per_1m": 0.5,
1592
  "currency": "USD",
1593
  "capabilities": [
1594
  "vision"
1595
  ],
1596
+ "context_window": 2000000,
1597
+ "hf_private": true
 
1598
  },
1599
  {
1600
+ "name": "xai/grok-code-fast-1",
1601
+ "type": "vision",
1602
  "input_price_per_1m": 0.2,
1603
+ "output_price_per_1m": 1.5,
1604
  "currency": "USD",
 
 
1605
  "capabilities": [
1606
+ "vision"
 
1607
  ],
1608
+ "context_window": 256000,
1609
+ "hf_private": true
1610
  },
1611
  {
1612
+ "name": "xai/grok-4-1-fast-non-reasoning",
1613
+ "type": "vision",
1614
  "input_price_per_1m": 0.2,
1615
+ "output_price_per_1m": 0.5,
1616
  "currency": "USD",
 
 
1617
  "capabilities": [
1618
+ "vision"
 
1619
  ],
1620
+ "context_window": 2000000,
1621
+ "hf_private": true
1622
  },
1623
  {
1624
  "name": "together/meta-llama/LlamaGuard-2-8b",
 
1630
  "size_b": 8
1631
  },
1632
  {
1633
+ "name": "openai-responses/gpt-5.4-nano",
1634
  "type": "vision",
1635
  "input_price_per_1m": 0.2,
1636
  "output_price_per_1m": 1.25,
 
1643
  "hf_private": true
1644
  },
1645
  {
1646
+ "name": "openai/gpt-5.4-nano",
1647
+ "type": "vision",
1648
  "input_price_per_1m": 0.2,
1649
+ "output_price_per_1m": 1.25,
1650
  "currency": "USD",
 
1651
  "capabilities": [
1652
  "vision",
 
1653
  "reasoning"
1654
  ],
1655
+ "context_window": 400000,
1656
  "hf_private": true
1657
  },
1658
  {
1659
+ "name": "deepinfra/zai-org/GLM-4.5-Air",
1660
+ "type": "chat",
1661
  "input_price_per_1m": 0.2,
1662
+ "output_price_per_1m": 1.1,
1663
  "currency": "USD",
1664
+ "context_window": 131072,
1665
  "capabilities": [
1666
+ "tools",
1667
+ "reasoning"
1668
  ],
1669
+ "hf_id": "zai-org/GLM-4.5-Air",
1670
+ "size_b": 110.5
1671
  },
1672
  {
1673
+ "name": "deepinfra/Qwen/Qwen3-235B-A22B",
1674
  "type": "chat",
1675
  "input_price_per_1m": 0.2,
1676
+ "output_price_per_1m": 0.6,
1677
  "currency": "USD",
1678
+ "context_window": 40000,
1679
+ "size_b": 235,
1680
  "capabilities": [
 
1681
  "tools",
1682
  "reasoning"
1683
  ],
1684
+ "hf_id": "Qwen/Qwen3-235B-A22B"
1685
  },
1686
  {
1687
+ "name": "novita/meta-llama/llama-4-maverick-17b-128e-instruct-fp8",
1688
+ "type": "chat",
1689
  "input_price_per_1m": 0.2,
1690
+ "output_price_per_1m": 0.85,
1691
  "currency": "USD",
1692
+ "context_window": 1048576,
1693
+ "size_b": 17,
1694
  "capabilities": [
1695
+ "vision",
1696
+ "tools"
1697
  ],
1698
+ "hf_id": "meta-llama/Llama-4-Maverick-17B-128E-Instruct"
 
1699
  },
1700
  {
1701
+ "name": "novita/qwen/qwen3-235b-a22b-fp8",
1702
+ "type": "chat",
1703
  "input_price_per_1m": 0.2,
1704
+ "output_price_per_1m": 0.8,
1705
  "currency": "USD",
1706
+ "context_window": 128000,
1707
+ "size_b": 235,
1708
  "capabilities": [
1709
+ "tools",
1710
+ "reasoning"
1711
  ],
1712
+ "hf_id": "Qwen/Qwen3-235B-A22B"
 
1713
  },
1714
  {
1715
  "name": "deepinfra/Qwen/Qwen2.5-72B-Instruct",
 
1724
  ],
1725
  "hf_id": "Qwen/Qwen2.5-VL-72B-Instruct"
1726
  },
 
 
 
 
 
 
 
 
 
 
 
 
 
1727
  {
1728
  "name": "deepinfra/meta-llama/Meta-Llama-3.1-70B-Instruct",
1729
  "type": "chat",
 
1746
  ],
1747
  "hf_id": "deepseek-ai/DeepSeek-R1-Distill-Llama-70B"
1748
  },
1749
+ {
1750
+ "name": "deepinfra/meta-llama/Llama-3.3-70B-Instruct",
1751
+ "type": "chat",
1752
+ "input_price_per_1m": 0.23,
1753
+ "output_price_per_1m": 0.4,
1754
+ "currency": "USD",
1755
+ "context_window": 131072,
1756
+ "size_b": 70,
1757
+ "capabilities": [
1758
+ "tools"
1759
+ ],
1760
+ "hf_id": "meta-llama/Llama-3.3-70B-Instruct"
1761
+ },
1762
  {
1763
  "name": "mistral/open-mistral-7b",
1764
  "type": "chat",
 
1769
  "size_b": 7
1770
  },
1771
  {
1772
+ "name": "google/gemini-3.1-flash-lite-preview",
1773
  "type": "vision",
1774
  "input_price_per_1m": 0.25,
1775
+ "output_price_per_1m": 1.5,
1776
  "currency": "USD",
1777
  "capabilities": [
1778
+ "vision"
 
1779
  ],
1780
+ "context_window": 1048576,
1781
  "hf_private": true
1782
  },
1783
  {
1784
+ "name": "openai-responses/gpt-5-mini",
1785
  "type": "vision",
1786
  "input_price_per_1m": 0.25,
1787
+ "output_price_per_1m": 2,
1788
  "currency": "USD",
1789
  "capabilities": [
1790
+ "vision",
1791
+ "reasoning"
1792
  ],
1793
+ "context_window": 400000,
1794
  "hf_private": true
1795
  },
1796
  {
 
1831
  "size_b": 685.4
1832
  },
1833
  {
1834
+ "name": "deepseek/deepseek-reasoner",
1835
  "type": "chat",
1836
  "input_price_per_1m": 0.28,
1837
  "output_price_per_1m": 0.42,
1838
  "currency": "USD",
1839
  "context_window": 128000,
1840
+ "hf_private": true
 
 
 
 
 
1841
  },
1842
  {
1843
+ "name": "deepseek/deepseek-chat",
1844
  "type": "chat",
1845
  "input_price_per_1m": 0.28,
1846
  "output_price_per_1m": 0.42,
1847
  "currency": "USD",
1848
  "context_window": 128000,
1849
+ "capabilities": [
1850
+ "tools"
1851
+ ],
1852
+ "hf_id": "deepseek-ai/DeepSeek-V3",
1853
+ "size_b": 671,
1854
+ "size_source": "manual"
1855
  },
1856
  {
1857
+ "name": "alibaba/qwen3-coder-flash",
1858
  "type": "vision",
1859
  "input_price_per_1m": 0.3,
1860
+ "output_price_per_1m": 1.5,
1861
  "currency": "USD",
1862
  "capabilities": [
1863
+ "vision"
 
1864
  ],
1865
  "context_window": 1048576,
1866
  "hf_private": true
 
1875
  "hf_private": true
1876
  },
1877
  {
1878
+ "name": "nebius/deepseek-ai/DeepSeek-V3.2",
1879
  "type": "chat",
1880
  "input_price_per_1m": 0.3,
1881
+ "output_price_per_1m": 0.45,
1882
  "currency": "USD",
 
1883
  "capabilities": [
 
1884
  "reasoning"
1885
  ],
1886
+ "context_window": 164000,
1887
+ "hf_id": "deepseek-ai/DeepSeek-V3.2",
1888
+ "size_b": 685.4
1889
  },
1890
  {
1891
+ "name": "nebius/minimax/minimax-m2.1",
1892
  "type": "chat",
1893
  "input_price_per_1m": 0.3,
1894
+ "output_price_per_1m": 1.2,
1895
  "currency": "USD",
1896
+ "capabilities": [
1897
+ "reasoning"
1898
+ ],
1899
+ "context_window": 200000,
1900
+ "hf_id": "MiniMaxAI/MiniMax-M2.1",
1901
+ "size_b": 228.7
1902
  },
1903
  {
1904
+ "name": "minimaxi/MiniMax-M2.5",
1905
  "type": "vision",
1906
  "input_price_per_1m": 0.3,
1907
+ "output_price_per_1m": 1.2,
1908
  "currency": "USD",
1909
  "capabilities": [
1910
+ "vision",
1911
+ "reasoning"
1912
  ],
1913
+ "context_window": 200000,
1914
+ "hf_id": "MiniMaxAI/MiniMax-M2.5",
1915
+ "size_b": 228.7
1916
  },
1917
  {
1918
+ "name": "minimaxi/MiniMax-M2.7",
1919
+ "type": "vision",
1920
  "input_price_per_1m": 0.3,
1921
  "output_price_per_1m": 1.2,
1922
  "currency": "USD",
1923
  "capabilities": [
1924
+ "vision",
1925
  "reasoning"
1926
  ],
1927
  "context_window": 200000,
1928
+ "hf_id": "MiniMaxAI/MiniMax-M2.7",
1929
+ "hf_private": true
1930
  },
1931
  {
1932
+ "name": "minimaxi/MiniMax-M2",
1933
  "type": "chat",
1934
  "input_price_per_1m": 0.3,
1935
+ "output_price_per_1m": 1.2,
1936
  "currency": "USD",
1937
+ "context_window": 200000,
1938
  "capabilities": [
1939
+ "tools",
1940
  "reasoning"
1941
  ],
1942
+ "hf_id": "MiniMaxAI/MiniMax-M2",
1943
+ "size_b": 228.7
 
1944
  },
1945
  {
1946
+ "name": "xai/grok-3-mini",
1947
  "type": "chat",
1948
  "input_price_per_1m": 0.3,
1949
+ "output_price_per_1m": 0.5,
1950
  "currency": "USD",
 
 
1951
  "capabilities": [
1952
  "reasoning"
1953
  ],
1954
+ "context_window": 131072,
1955
+ "hf_private": true
1956
  },
1957
  {
1958
+ "name": "google/gemini-2.5-flash",
1959
  "type": "vision",
1960
  "input_price_per_1m": 0.3,
1961
+ "output_price_per_1m": 2.5,
1962
  "currency": "USD",
1963
  "capabilities": [
1964
  "vision",
1965
  "reasoning"
1966
  ],
1967
+ "context_window": 1048576,
 
1968
  "hf_private": true
1969
  },
1970
  {
1971
+ "name": "together/Qwen/Qwen2.5-7B-Instruct-Turbo",
1972
+ "type": "chat",
1973
+ "input_price_per_1m": 0.3,
1974
+ "output_price_per_1m": 0.3,
1975
+ "currency": "USD",
1976
+ "context_window": 32768,
1977
+ "size_b": 7
1978
+ },
1979
+ {
1980
+ "name": "coding/gemini-2.5-flash",
1981
  "type": "vision",
1982
  "input_price_per_1m": 0.3,
1983
+ "output_price_per_1m": 2.5,
1984
  "currency": "USD",
1985
  "capabilities": [
1986
  "vision",
1987
  "reasoning"
1988
  ],
1989
+ "context_window": 1048576,
 
1990
  "hf_private": true
1991
  },
1992
  {
1993
+ "name": "deepinfra/deepseek-ai/DeepSeek-V3.1",
1994
+ "type": "chat",
1995
  "input_price_per_1m": 0.3,
1996
+ "output_price_per_1m": 1,
1997
  "currency": "USD",
1998
+ "context_window": 163840,
1999
  "capabilities": [
2000
+ "tools",
2001
  "reasoning"
2002
  ],
2003
+ "hf_id": "deepseek-ai/DeepSeek-V3.1-Terminus",
2004
+ "size_b": 671
2005
+ },
2006
+ {
2007
+ "name": "parasail/parasail-gemma3-27b-it",
2008
+ "type": "chat",
2009
+ "input_price_per_1m": 0.3,
2010
+ "output_price_per_1m": 0.5,
2011
+ "currency": "USD",
2012
+ "context_window": 128000,
2013
+ "size_b": 27
2014
  },
2015
  {
2016
  "name": "vertex/gemini-2.5-flash",
 
2039
  "hf_private": true
2040
  },
2041
  {
2042
+ "name": "novita/deepseek/deepseek-r1-distill-qwen-32b",
2043
  "type": "chat",
2044
  "input_price_per_1m": 0.3,
2045
+ "output_price_per_1m": 0.3,
2046
  "currency": "USD",
2047
+ "context_window": 12800,
2048
+ "size_b": 32,
2049
  "capabilities": [
 
2050
  "reasoning"
2051
  ],
2052
+ "hf_id": "deepseek-ai/DeepSeek-R1-Distill-Qwen-32B"
 
2053
  },
2054
  {
2055
+ "name": "novita/MiniMax-M2.7",
2056
  "type": "vision",
2057
  "input_price_per_1m": 0.3,
2058
  "output_price_per_1m": 1.2,
 
2066
  "hf_private": true
2067
  },
2068
  {
2069
+ "name": "novita/minimax-m2.7",
2070
  "type": "vision",
2071
  "input_price_per_1m": 0.3,
2072
  "output_price_per_1m": 1.2,
 
2076
  "reasoning"
2077
  ],
2078
  "context_window": 200000,
2079
+ "hf_id": "MiniMaxAI/MiniMax-M2.7",
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2080
  "hf_private": true
2081
  },
2082
  {
 
2104
  {
2105
  "name": "novita/meta-llama/llama-3.3-70b-instruct",
2106
  "type": "chat",
2107
+ "input_price_per_1m": 0.39,
2108
+ "output_price_per_1m": 0.39,
2109
+ "currency": "USD",
2110
+ "context_window": 131072,
2111
+ "size_b": 70,
2112
+ "capabilities": [
2113
+ "tools"
2114
+ ],
2115
+ "hf_id": "meta-llama/Llama-3.3-70B-Instruct"
2116
+ },
2117
+ {
2118
+ "name": "alibaba/qwen-plus",
2119
+ "type": "chat",
2120
+ "input_price_per_1m": 0.4,
2121
+ "output_price_per_1m": 1.2,
2122
  "currency": "USD",
2123
  "context_window": 131072,
 
2124
  "capabilities": [
2125
  "tools"
2126
  ],
2127
+ "hf_private": true
2128
  },
2129
  {
2130
  "name": "mistral/mistral-medium-latest",
 
2156
  ],
2157
  "hf_private": true
2158
  },
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2159
  {
2160
  "name": "nebius/zai-org/GLM-4.7",
2161
  "type": "chat",
 
2183
  "hf_id": "Qwen/Qwen3-Coder-480B-A35B-Instruct"
2184
  },
2185
  {
2186
+ "name": "openai-responses/gpt-4.1-mini",
2187
+ "type": "vision",
2188
  "input_price_per_1m": 0.4,
2189
+ "output_price_per_1m": 1.6,
2190
  "currency": "USD",
 
2191
  "capabilities": [
2192
+ "vision"
 
2193
  ],
2194
+ "context_window": 1047576,
 
 
 
 
 
 
 
 
 
2195
  "hf_private": true
2196
  },
2197
  {
2198
+ "name": "openai/gpt-4.1-mini",
2199
  "type": "vision",
2200
  "input_price_per_1m": 0.4,
2201
  "output_price_per_1m": 1.6,
 
2206
  "context_window": 1047576,
2207
  "hf_private": true
2208
  },
2209
+ {
2210
+ "name": "deepinfra/Qwen/Qwen3-Coder-480B-A35B-Instruct",
2211
+ "type": "chat",
2212
+ "input_price_per_1m": 0.4,
2213
+ "output_price_per_1m": 1.6,
2214
+ "currency": "USD",
2215
+ "context_window": 262144,
2216
+ "size_b": 480,
2217
+ "capabilities": [
2218
+ "tools"
2219
+ ],
2220
+ "hf_id": "Qwen/Qwen3-Coder-480B-A35B-Instruct"
2221
+ },
2222
  {
2223
  "name": "azure/openai-responses/gpt-4.1-mini",
2224
  "type": "vision",
 
2232
  "hf_private": true
2233
  },
2234
  {
2235
+ "name": "azure/gpt-4.1-mini",
2236
  "type": "vision",
2237
  "input_price_per_1m": 0.4,
2238
  "output_price_per_1m": 1.6,
 
2244
  "hf_private": true
2245
  },
2246
  {
2247
+ "name": "novita/deepseek/deepseek-v3-turbo",
2248
  "type": "chat",
2249
+ "input_price_per_1m": 0.4,
2250
+ "output_price_per_1m": 1.3,
2251
  "currency": "USD",
2252
+ "context_window": 128000,
2253
+ "hf_private": true
2254
+ },
2255
+ {
2256
+ "name": "novita/deepseek/deepseek-v3-0324",
2257
+ "type": "chat",
2258
+ "input_price_per_1m": 0.4,
2259
+ "output_price_per_1m": 1.3,
2260
+ "currency": "USD",
2261
+ "context_window": 128000,
2262
  "capabilities": [
2263
+ "tools",
2264
+ "reasoning"
2265
  ],
2266
+ "size_b": 685,
2267
+ "hf_id": "deepseek-ai/DeepSeek-V3-0324"
2268
  },
2269
  {
2270
+ "name": "mistral/mistral-large-latest",
2271
  "type": "chat",
2272
  "input_price_per_1m": 0.5,
2273
+ "output_price_per_1m": 1.5,
2274
  "currency": "USD",
2275
+ "context_window": 131072,
2276
  "capabilities": [
2277
  "tools"
2278
  ],
2279
+ "hf_private": true
 
2280
  },
2281
  {
2282
  "name": "nebius/moonshotai/kimi-k2.5",
 
2292
  "hf_id": "moonshotai/Kimi-K2.5",
2293
  "size_b": 1058.6
2294
  },
2295
+ {
2296
+ "name": "nebius/moonshotai/Kimi-K2-Instruct",
2297
+ "type": "chat",
2298
+ "input_price_per_1m": 0.5,
2299
+ "output_price_per_1m": 2.4,
2300
+ "currency": "USD",
2301
+ "context_window": 131000,
2302
+ "capabilities": [
2303
+ "tools"
2304
+ ],
2305
+ "hf_id": "moonshotai/Kimi-K2-Instruct",
2306
+ "size_b": 1026.5
2307
+ },
2308
  {
2309
  "name": "nebius/deepseek-ai/DeepSeek-V3-0324",
2310
  "type": "chat",
 
2346
  "hf_private": true
2347
  },
2348
  {
2349
+ "name": "vertex/gemini-3-flash-preview",
2350
  "type": "vision",
2351
  "input_price_per_1m": 0.5,
2352
+ "output_price_per_1m": 3,
2353
  "currency": "USD",
2354
  "capabilities": [
2355
  "vision",
2356
  "reasoning"
2357
  ],
2358
+ "context_window": 1048576,
2359
  "hf_private": true
2360
  },
2361
  {
2362
+ "name": "vertex/gemini-3.1-flash-image-preview",
2363
  "type": "vision",
2364
  "input_price_per_1m": 0.5,
2365
+ "output_price_per_1m": 2,
2366
  "currency": "USD",
2367
  "capabilities": [
2368
  "vision",
2369
  "reasoning"
2370
  ],
2371
+ "context_window": 131072,
2372
  "hf_private": true
2373
  },
2374
  {
 
2423
  "hf_id": "moonshotai/Kimi-K2-Instruct",
2424
  "size_b": 1026.5
2425
  },
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2426
  {
2427
  "name": "zai/GLM-4.7",
2428
  "type": "chat",
 
2491
  "size_b": 358.3
2492
  },
2493
  {
2494
+ "name": "minimaxi/MiniMax-M2.7-highspeed",
2495
+ "type": "vision",
2496
+ "input_price_per_1m": 0.6,
2497
+ "output_price_per_1m": 2.4,
2498
+ "currency": "USD",
2499
+ "capabilities": [
2500
+ "vision",
2501
+ "reasoning"
2502
+ ],
2503
+ "context_window": 200000,
2504
+ "hf_id": "MiniMaxAI/MiniMax-M2.7",
2505
+ "hf_private": true
2506
+ },
2507
+ {
2508
+ "name": "minimaxi/MiniMax-M2.5-highspeed",
2509
+ "type": "vision",
2510
+ "input_price_per_1m": 0.6,
2511
+ "output_price_per_1m": 2.4,
2512
+ "currency": "USD",
2513
+ "capabilities": [
2514
+ "vision",
2515
+ "reasoning"
2516
+ ],
2517
+ "context_window": 200000,
2518
+ "hf_id": "MiniMaxAI/MiniMax-M2.5",
2519
+ "size_b": 228.7
2520
+ },
2521
+ {
2522
+ "name": "deepinfra/zai-org/GLM-4.5",
2523
  "type": "chat",
2524
  "input_price_per_1m": 0.6,
2525
  "output_price_per_1m": 2.2,
 
2547
  "size_b": 356.8
2548
  },
2549
  {
2550
+ "name": "novita/zai-org/glm-4.5",
2551
+ "type": "chat",
2552
  "input_price_per_1m": 0.6,
2553
+ "output_price_per_1m": 2.2,
2554
  "currency": "USD",
2555
+ "context_window": 131072,
2556
  "capabilities": [
2557
+ "tools",
2558
  "reasoning"
2559
  ],
2560
+ "hf_id": "zai-org/GLM-4.5",
2561
+ "size_b": 358.3
 
2562
  },
2563
  {
2564
  "name": "moonshot/kimi-k2-thinking-turbo",
 
2614
  "size_b": 1026.5
2615
  },
2616
  {
2617
+ "name": "moonshot/kimi-k2.5",
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2618
  "type": "vision",
2619
  "input_price_per_1m": 0.6,
2620
+ "output_price_per_1m": 3,
2621
  "currency": "USD",
2622
  "capabilities": [
2623
  "vision",
2624
  "reasoning"
2625
  ],
2626
+ "context_window": 262144,
2627
+ "hf_id": "moonshotai/Kimi-K2.5",
2628
+ "size_b": 1058.6
2629
  },
2630
  {
2631
  "name": "novita/microsoft/wizardlm-2-8x22b",
 
2712
  "context_window": 400000,
2713
  "hf_private": true
2714
  },
 
 
 
 
 
 
 
 
 
2715
  {
2716
  "name": "nebius/deepseek-ai/DeepSeek-R1-0528",
2717
  "type": "chat",
 
2726
  "hf_id": "deepseek-ai/DeepSeek-R1-0528",
2727
  "size_b": 684.5
2728
  },
2729
+ {
2730
+ "name": "deepinfra/meta-llama/Meta-Llama-3.1-405B-Instruct",
2731
+ "type": "chat",
2732
+ "input_price_per_1m": 0.8,
2733
+ "output_price_per_1m": 0.8,
2734
+ "currency": "USD",
2735
+ "context_window": 130815,
2736
+ "size_b": 405
2737
+ },
2738
  {
2739
  "name": "novita/qwen/qwen2.5-vl-72b-instruct",
2740
  "type": "chat",
 
2758
  "size_b": 70,
2759
  "capabilities": [
2760
  "reasoning"
2761
+ ],
2762
+ "hf_id": "deepseek-ai/DeepSeek-R1-Distill-Llama-70B"
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2763
  },
2764
  {
2765
  "name": "deepinfra/deepseek-ai/DeepSeek-R1",
 
2776
  "size_b": 671,
2777
  "size_source": "manual"
2778
  },
2779
+ {
2780
+ "name": "deepinfra/deepseek-ai/DeepSeek-V3",
2781
+ "type": "chat",
2782
+ "input_price_per_1m": 0.85,
2783
+ "output_price_per_1m": 0.9,
2784
+ "currency": "USD",
2785
+ "context_window": 128000,
2786
+ "capabilities": [
2787
+ "tools",
2788
+ "reasoning"
2789
+ ],
2790
+ "hf_id": "deepseek-ai/DeepSeek-V3.1-Terminus",
2791
+ "size_b": 671
2792
+ },
2793
  {
2794
  "name": "alibaba/qwen3-max",
2795
  "type": "vision",
 
2803
  "hf_private": true
2804
  },
2805
  {
2806
+ "name": "together/meta-llama/Llama-3.3-70B-Instruct-Turbo",
2807
  "type": "chat",
2808
  "input_price_per_1m": 0.88,
2809
  "output_price_per_1m": 0.88,
2810
  "currency": "USD",
2811
  "context_window": 131072,
2812
+ "size_b": 70,
2813
+ "capabilities": [
2814
+ "tools"
2815
+ ],
2816
+ "hf_id": "meta-llama/Llama-3.3-70B-Instruct"
2817
  },
2818
  {
2819
+ "name": "together/meta-llama/Meta-Llama-3.1-70B-Instruct-Turbo",
2820
  "type": "chat",
2821
  "input_price_per_1m": 0.88,
2822
  "output_price_per_1m": 0.88,
2823
  "currency": "USD",
2824
  "context_window": 131072,
2825
+ "size_b": 70
 
 
 
 
2826
  },
2827
  {
2828
  "name": "novita/deepseek/deepseek_v3",
 
2847
  "context_window": 131072,
2848
  "hf_private": true
2849
  },
 
 
 
 
 
 
 
 
 
 
 
 
2850
  {
2851
  "name": "zai/GLM-5",
2852
  "type": "vision",
 
2874
  "hf_private": true
2875
  },
2876
  {
2877
+ "name": "perplexity/sonar",
2878
+ "type": "chat",
2879
  "input_price_per_1m": 1,
2880
+ "output_price_per_1m": 1,
2881
  "currency": "USD",
2882
+ "context_window": 131072,
2883
  "capabilities": [
2884
  "vision"
2885
  ],
 
2886
  "hf_private": true
2887
  },
2888
  {
2889
+ "name": "anthropic/claude-haiku-4-5",
2890
+ "type": "vision",
2891
  "input_price_per_1m": 1,
2892
+ "output_price_per_1m": 5,
2893
  "currency": "USD",
2894
  "capabilities": [
2895
+ "vision"
2896
  ],
2897
+ "context_window": 200000,
2898
+ "hf_private": true
 
2899
  },
2900
  {
2901
+ "name": "bedrock/claude-haiku-4-5",
2902
+ "type": "vision",
2903
  "input_price_per_1m": 1,
2904
+ "output_price_per_1m": 5,
2905
  "currency": "USD",
 
2906
  "capabilities": [
2907
  "vision"
2908
  ],
2909
+ "context_window": 200000,
2910
  "hf_private": true
2911
  },
2912
  {
2913
+ "name": "openai/o3",
2914
  "type": "vision",
2915
  "input_price_per_1m": 1,
2916
+ "output_price_per_1m": 4,
2917
  "currency": "USD",
2918
  "capabilities": [
2919
+ "vision",
2920
+ "reasoning"
2921
  ],
2922
  "context_window": 200000,
2923
  "hf_private": true
2924
  },
2925
  {
2926
+ "name": "groq/moonshotai/kimi-k2-instruct",
2927
  "type": "chat",
2928
  "input_price_per_1m": 1,
2929
  "output_price_per_1m": 3,
2930
  "currency": "USD",
2931
+ "context_window": 131072,
2932
  "capabilities": [
2933
  "tools"
2934
  ],
 
2936
  "size_b": 1026.5
2937
  },
2938
  {
2939
+ "name": "groq/moonshotai/Kimi-K2-Instruct-0905",
2940
  "type": "chat",
2941
  "input_price_per_1m": 1,
2942
  "output_price_per_1m": 3,
2943
  "currency": "USD",
2944
+ "context_window": 256000,
2945
  "capabilities": [
2946
  "tools"
2947
  ],
 
2949
  "size_b": 1026.5
2950
  },
2951
  {
2952
+ "name": "vertex/claude-haiku-4-5",
2953
  "type": "vision",
2954
  "input_price_per_1m": 1,
2955
+ "output_price_per_1m": 5,
2956
  "currency": "USD",
2957
  "capabilities": [
2958
+ "vision"
 
2959
  ],
2960
  "context_window": 200000,
2961
  "hf_private": true
2962
  },
2963
+ {
2964
+ "name": "novita/GLM-5",
2965
+ "type": "chat",
2966
+ "input_price_per_1m": 1,
2967
+ "output_price_per_1m": 3.2,
2968
+ "currency": "USD",
2969
+ "capabilities": [
2970
+ "reasoning"
2971
+ ],
2972
+ "context_window": 202800,
2973
+ "hf_id": "zai-org/GLM-5",
2974
+ "size_b": 753.9
2975
+ },
2976
  {
2977
  "name": "openai-responses/o3-mini",
2978
  "type": "chat",
 
2998
  "hf_private": true
2999
  },
3000
  {
3001
+ "name": "openai/o3-mini",
3002
  "type": "chat",
3003
  "input_price_per_1m": 1.1,
3004
  "output_price_per_1m": 4.4,
 
3010
  "hf_private": true
3011
  },
3012
  {
3013
+ "name": "azure/o4-mini",
3014
  "type": "chat",
3015
  "input_price_per_1m": 1.1,
3016
  "output_price_per_1m": 4.4,
 
3021
  "context_window": 200000,
3022
  "hf_private": true
3023
  },
3024
+ {
3025
+ "name": "together/Qwen/Qwen2.5-72B-Instruct-Turbo",
3026
+ "type": "chat",
3027
+ "input_price_per_1m": 1.2,
3028
+ "output_price_per_1m": 1.2,
3029
+ "currency": "USD",
3030
+ "context_window": 32768,
3031
+ "size_b": 72
3032
+ },
3033
  {
3034
  "name": "moonshot/kimi-k2-turbo-preview",
3035
  "type": "chat",
 
3044
  "size_b": 1026.5
3045
  },
3046
  {
3047
+ "name": "google/gemini-2.5-pro",
 
 
 
 
 
 
 
 
 
3048
  "type": "vision",
3049
  "input_price_per_1m": 1.25,
3050
  "output_price_per_1m": 10,
 
3057
  "hf_private": true
3058
  },
3059
  {
3060
+ "name": "together/deepseek-ai/DeepSeek-V3",
3061
+ "type": "chat",
3062
  "input_price_per_1m": 1.25,
3063
+ "output_price_per_1m": 1.25,
3064
  "currency": "USD",
3065
+ "context_window": 131072,
3066
  "capabilities": [
3067
+ "tools",
3068
  "reasoning"
3069
  ],
3070
+ "hf_id": "deepseek-ai/DeepSeek-V3.1-Terminus",
3071
+ "size_b": 671
3072
  },
3073
  {
3074
+ "name": "openai-responses/gpt-5.1-codex",
3075
  "type": "vision",
3076
  "input_price_per_1m": 1.25,
3077
  "output_price_per_1m": 10,
 
3084
  "hf_private": true
3085
  },
3086
  {
3087
+ "name": "openai-responses/gpt-5.1",
3088
  "type": "vision",
3089
  "input_price_per_1m": 1.25,
3090
  "output_price_per_1m": 10,
 
3110
  "hf_private": true
3111
  },
3112
  {
3113
+ "name": "openai-responses/gpt-5",
3114
  "type": "vision",
3115
  "input_price_per_1m": 1.25,
3116
  "output_price_per_1m": 10,
 
3119
  "vision",
3120
  "reasoning"
3121
  ],
3122
+ "context_window": 400000,
3123
  "hf_private": true
3124
  },
3125
  {
3126
+ "name": "openai/gpt-5-chat",
3127
  "type": "vision",
3128
  "input_price_per_1m": 1.25,
3129
  "output_price_per_1m": 10,
 
3132
  "vision",
3133
  "reasoning"
3134
  ],
3135
+ "context_window": 128000,
3136
  "hf_private": true
3137
  },
3138
  {
3139
+ "name": "openai/gpt-5.1-chat",
3140
+ "type": "vision",
3141
  "input_price_per_1m": 1.25,
3142
  "output_price_per_1m": 10,
3143
  "currency": "USD",
3144
  "capabilities": [
3145
+ "vision",
3146
  "reasoning"
3147
  ],
3148
+ "context_window": 128000,
3149
  "hf_private": true
3150
  },
3151
  {
3152
+ "name": "openai/gpt-5.1",
3153
+ "type": "vision",
3154
  "input_price_per_1m": 1.25,
3155
  "output_price_per_1m": 10,
3156
  "currency": "USD",
3157
  "capabilities": [
3158
+ "vision",
3159
  "reasoning"
3160
  ],
3161
+ "context_window": 400000,
3162
  "hf_private": true
3163
  },
3164
  {
3165
+ "name": "coding/gemini-2.5-pro",
3166
+ "type": "vision",
3167
  "input_price_per_1m": 1.25,
3168
+ "output_price_per_1m": 10,
3169
  "currency": "USD",
 
3170
  "capabilities": [
3171
+ "vision",
3172
  "reasoning"
3173
  ],
3174
+ "context_window": 1048576,
3175
+ "hf_private": true
3176
  },
3177
  {
3178
+ "name": "vertex/gemini-2.5-pro",
3179
  "type": "vision",
3180
  "input_price_per_1m": 1.25,
3181
  "output_price_per_1m": 10,
 
3184
  "vision",
3185
  "reasoning"
3186
  ],
3187
+ "context_window": 1048576,
3188
  "hf_private": true
3189
  },
3190
  {
3191
+ "name": "azure/gpt-5",
3192
+ "type": "chat",
3193
  "input_price_per_1m": 1.25,
3194
  "output_price_per_1m": 10,
3195
  "currency": "USD",
3196
  "capabilities": [
 
3197
  "reasoning"
3198
  ],
3199
+ "context_window": 200000,
3200
  "hf_private": true
3201
  },
3202
  {
3203
+ "name": "azure/gpt-5.1",
3204
+ "type": "chat",
3205
  "input_price_per_1m": 1.25,
3206
  "output_price_per_1m": 10,
3207
  "currency": "USD",
3208
  "capabilities": [
 
3209
  "reasoning"
3210
  ],
3211
+ "context_window": 200000,
3212
  "hf_private": true
3213
  },
3214
  {
3215
+ "name": "novita/sao10k/l3-70b-euryale-v2.1",
3216
  "type": "chat",
3217
  "input_price_per_1m": 1.48,
3218
  "output_price_per_1m": 1.48,
 
3221
  "size_b": 70
3222
  },
3223
  {
3224
+ "name": "novita/sao10k/l31-70b-euryale-v2.2",
3225
  "type": "chat",
3226
  "input_price_per_1m": 1.48,
3227
  "output_price_per_1m": 1.48,
 
3242
  "hf_private": true
3243
  },
3244
  {
3245
+ "name": "openai-responses/gpt-5.3-chat",
3246
  "type": "vision",
3247
  "input_price_per_1m": 1.75,
3248
  "output_price_per_1m": 14,
 
3251
  "vision",
3252
  "reasoning"
3253
  ],
3254
+ "context_window": 128000,
3255
  "hf_private": true
3256
  },
3257
  {
3258
+ "name": "openai-responses/gpt-5.3-codex",
3259
  "type": "vision",
3260
  "input_price_per_1m": 1.75,
3261
  "output_price_per_1m": 14,
 
3264
  "vision",
3265
  "reasoning"
3266
  ],
3267
+ "context_window": 400000,
3268
  "hf_private": true
3269
  },
3270
  {
 
3281
  "hf_private": true
3282
  },
3283
  {
3284
+ "name": "openai-responses/gpt-5.2",
3285
  "type": "vision",
3286
  "input_price_per_1m": 1.75,
3287
  "output_price_per_1m": 14,
 
3294
  "hf_private": true
3295
  },
3296
  {
3297
+ "name": "openai/gpt-5.2",
3298
  "type": "vision",
3299
  "input_price_per_1m": 1.75,
3300
  "output_price_per_1m": 14,
 
3307
  "hf_private": true
3308
  },
3309
  {
3310
+ "name": "openai/gpt-5.2-chat",
3311
+ "type": "vision",
3312
  "input_price_per_1m": 1.75,
3313
  "output_price_per_1m": 14,
3314
  "currency": "USD",
3315
  "capabilities": [
3316
+ "vision",
3317
  "reasoning"
3318
  ],
3319
+ "context_window": 128000,
3320
  "hf_private": true
3321
  },
3322
  {
3323
+ "name": "openai/gpt-5.3-chat",
3324
  "type": "vision",
3325
  "input_price_per_1m": 1.75,
3326
  "output_price_per_1m": 14,
 
3329
  "vision",
3330
  "reasoning"
3331
  ],
3332
+ "context_window": 128000,
3333
  "hf_private": true
3334
  },
3335
  {
3336
+ "name": "azure/gpt-5.2",
3337
+ "type": "chat",
3338
  "input_price_per_1m": 1.75,
3339
  "output_price_per_1m": 14,
3340
  "currency": "USD",
3341
  "capabilities": [
 
3342
  "reasoning"
3343
  ],
3344
+ "context_window": 200000,
3345
  "hf_private": true
3346
  },
3347
  {
3348
+ "name": "azure/openai-responses/gpt-5.2-codex",
3349
  "type": "vision",
3350
  "input_price_per_1m": 1.75,
3351
  "output_price_per_1m": 14,
 
3354
  "vision",
3355
  "reasoning"
3356
  ],
3357
+ "context_window": 400000,
3358
  "hf_private": true
3359
  },
3360
  {
3361
+ "name": "azure/gpt-5.3-codex",
3362
  "type": "vision",
3363
  "input_price_per_1m": 1.75,
3364
  "output_price_per_1m": 14,
 
3367
  "vision",
3368
  "reasoning"
3369
  ],
3370
+ "context_window": 400000,
3371
  "hf_private": true
3372
  },
3373
  {
3374
+ "name": "perplexity/sonar-reasoning-pro",
3375
  "type": "chat",
3376
  "input_price_per_1m": 2,
3377
+ "output_price_per_1m": 8,
3378
  "currency": "USD",
3379
  "context_window": 131072,
3380
+ "capabilities": [
3381
+ "vision",
3382
+ "reasoning"
3383
+ ],
3384
  "hf_private": true
3385
  },
3386
  {
3387
+ "name": "mistral/pixtral-large-latest",
3388
+ "type": "chat",
3389
  "input_price_per_1m": 2,
3390
+ "output_price_per_1m": 5,
3391
  "currency": "USD",
3392
+ "context_window": 131072,
 
 
 
3393
  "hf_private": true
3394
  },
3395
  {
 
3402
  "hf_private": true
3403
  },
3404
  {
3405
+ "name": "xai/grok-2-1212",
3406
  "type": "chat",
3407
  "input_price_per_1m": 2,
3408
+ "output_price_per_1m": 10,
3409
  "currency": "USD",
3410
  "context_window": 131072,
 
 
 
 
3411
  "hf_private": true
3412
  },
3413
  {
3414
+ "name": "xai/grok-4.2-beta",
3415
  "type": "vision",
3416
  "input_price_per_1m": 2,
3417
+ "output_price_per_1m": 6,
3418
  "currency": "USD",
3419
  "capabilities": [
3420
  "vision",
3421
  "reasoning"
3422
  ],
3423
+ "context_window": 2000000,
3424
  "hf_private": true
3425
  },
3426
  {
3427
+ "name": "google/gemini-3.1-pro-preview",
3428
  "type": "vision",
3429
  "input_price_per_1m": 2,
3430
  "output_price_per_1m": 12,
 
3437
  "hf_private": true
3438
  },
3439
  {
3440
+ "name": "google/gemini-3-pro-preview",
3441
  "type": "vision",
3442
  "input_price_per_1m": 2,
3443
  "output_price_per_1m": 12,
 
3450
  "hf_private": true
3451
  },
3452
  {
3453
+ "name": "openai-responses/gpt-4.1",
3454
  "type": "vision",
3455
  "input_price_per_1m": 2,
3456
+ "output_price_per_1m": 8,
3457
  "currency": "USD",
3458
  "capabilities": [
3459
+ "vision"
 
3460
  ],
3461
+ "context_window": 1047576,
3462
  "hf_private": true
3463
  },
3464
  {
3465
+ "name": "openai/o4-mini-deep-research",
3466
  "type": "vision",
3467
  "input_price_per_1m": 2,
3468
+ "output_price_per_1m": 8,
3469
  "currency": "USD",
3470
  "capabilities": [
3471
  "vision",
3472
  "reasoning"
3473
  ],
3474
+ "context_window": 100000,
3475
  "hf_private": true
3476
  },
3477
  {
3478
+ "name": "openai/gpt-4.1",
3479
  "type": "vision",
3480
  "input_price_per_1m": 2,
3481
  "output_price_per_1m": 8,
 
3487
  "hf_private": true
3488
  },
3489
  {
3490
+ "name": "vertex/gemini-3.1-pro-preview",
3491
  "type": "vision",
3492
  "input_price_per_1m": 2,
3493
+ "output_price_per_1m": 12,
3494
  "currency": "USD",
3495
  "capabilities": [
3496
+ "vision",
3497
+ "reasoning"
3498
  ],
3499
+ "context_window": 1048576,
3500
  "hf_private": true
3501
  },
3502
  {
3503
+ "name": "vertex/gemini-3-pro-image-preview",
3504
  "type": "vision",
3505
  "input_price_per_1m": 2,
3506
+ "output_price_per_1m": 12,
3507
  "currency": "USD",
3508
  "capabilities": [
3509
+ "vision",
3510
+ "reasoning"
3511
  ],
3512
+ "context_window": 1048576,
3513
  "hf_private": true
3514
  },
3515
  {
3516
+ "name": "vertex/gemini-3-pro-preview",
3517
  "type": "vision",
3518
  "input_price_per_1m": 2,
3519
+ "output_price_per_1m": 12,
3520
  "currency": "USD",
3521
  "capabilities": [
3522
  "vision",
3523
  "reasoning"
3524
  ],
3525
+ "context_window": 1048576,
3526
  "hf_private": true
3527
  },
3528
  {
3529
+ "name": "azure/openai-responses/gpt-4.1",
3530
  "type": "vision",
3531
  "input_price_per_1m": 2,
3532
+ "output_price_per_1m": 8,
3533
  "currency": "USD",
3534
  "capabilities": [
3535
+ "vision"
 
3536
  ],
3537
+ "context_window": 1047576,
3538
  "hf_private": true
3539
  },
3540
  {
3541
+ "name": "azure/gpt-4.1",
3542
+ "type": "vision",
3543
  "input_price_per_1m": 2,
3544
+ "output_price_per_1m": 8,
3545
  "currency": "USD",
3546
+ "capabilities": [
3547
+ "vision"
3548
+ ],
3549
+ "context_window": 1047576,
3550
  "hf_private": true
3551
  },
3552
  {
 
3563
  "hf_private": true
3564
  },
3565
  {
3566
+ "name": "openai/gpt-4o-2024-08-06",
3567
  "type": "vision",
3568
  "input_price_per_1m": 2.5,
3569
+ "output_price_per_1m": 10,
3570
  "currency": "USD",
3571
  "capabilities": [
3572
+ "vision"
 
3573
  ],
3574
+ "context_window": 128000,
3575
  "hf_private": true
3576
  },
3577
  {
3578
+ "name": "openai/gpt-4o-2024-05-13",
3579
  "type": "vision",
3580
  "input_price_per_1m": 2.5,
3581
  "output_price_per_1m": 10,
 
3587
  "hf_private": true
3588
  },
3589
  {
3590
+ "name": "openai/gpt-5.4",
3591
  "type": "vision",
3592
  "input_price_per_1m": 2.5,
3593
+ "output_price_per_1m": 15,
3594
  "currency": "USD",
3595
  "capabilities": [
3596
+ "vision",
3597
+ "reasoning"
3598
  ],
3599
+ "context_window": 1050000,
3600
  "hf_private": true
3601
  },
3602
  {
3603
+ "name": "openai/gpt-4o",
3604
  "type": "vision",
3605
  "input_price_per_1m": 2.5,
3606
  "output_price_per_1m": 10,
 
3624
  "hf_private": true
3625
  },
3626
  {
3627
+ "name": "azure/openai-responses/gpt-5.4",
3628
  "type": "vision",
3629
  "input_price_per_1m": 2.5,
3630
  "output_price_per_1m": 15,
 
3637
  "hf_private": true
3638
  },
3639
  {
3640
+ "name": "perplexity/sonar-pro",
3641
+ "type": "chat",
3642
  "input_price_per_1m": 3,
3643
  "output_price_per_1m": 15,
3644
  "currency": "USD",
3645
+ "context_window": 204800,
3646
  "capabilities": [
3647
+ "vision"
 
3648
  ],
 
3649
  "hf_private": true
3650
  },
3651
  {
3652
+ "name": "anthropic/claude-sonnet-4-5",
3653
  "type": "vision",
3654
  "input_price_per_1m": 3,
3655
  "output_price_per_1m": 15,
 
3662
  "hf_private": true
3663
  },
3664
  {
3665
+ "name": "anthropic/claude-sonnet-4",
3666
  "type": "vision",
3667
  "input_price_per_1m": 3,
3668
  "output_price_per_1m": 15,
 
3675
  "hf_private": true
3676
  },
3677
  {
3678
+ "name": "anthropic/claude-sonnet-4-6",
3679
  "type": "vision",
3680
  "input_price_per_1m": 3,
3681
  "output_price_per_1m": 15,
 
3688
  "hf_private": true
3689
  },
3690
  {
3691
+ "name": "xai/grok-4",
3692
  "type": "vision",
3693
  "input_price_per_1m": 3,
3694
  "output_price_per_1m": 15,
3695
  "currency": "USD",
3696
  "capabilities": [
3697
+ "vision"
 
3698
  ],
3699
+ "context_window": 256000,
3700
  "hf_private": true
3701
  },
3702
  {
 
3713
  "hf_private": true
3714
  },
3715
  {
3716
+ "name": "bedrock/claude-sonnet-4",
3717
  "type": "vision",
3718
  "input_price_per_1m": 3,
3719
  "output_price_per_1m": 15,
 
3722
  "vision",
3723
  "reasoning"
3724
  ],
3725
+ "context_window": 200000,
3726
  "hf_private": true
3727
  },
3728
  {
 
3739
  "hf_private": true
3740
  },
3741
  {
3742
+ "name": "bedrock/claude-sonnet-4-6",
3743
+ "type": "vision",
3744
  "input_price_per_1m": 3,
3745
  "output_price_per_1m": 15,
3746
  "currency": "USD",
 
3747
  "capabilities": [
3748
+ "vision",
3749
+ "reasoning"
3750
  ],
3751
+ "context_window": 1000000,
3752
  "hf_private": true
3753
  },
3754
  {
3755
+ "name": "together/deepseek-ai/DeepSeek-R1",
3756
+ "type": "chat",
3757
  "input_price_per_1m": 3,
3758
+ "output_price_per_1m": 7,
3759
  "currency": "USD",
3760
+ "context_window": 64000,
3761
  "capabilities": [
3762
+ "tools",
3763
  "reasoning"
3764
  ],
3765
+ "hf_id": "deepseek-ai/DeepSeek-R1",
3766
+ "size_b": 671,
3767
+ "size_source": "manual"
3768
  },
3769
  {
3770
+ "name": "coding/claude-sonnet-4-20250514",
3771
  "type": "vision",
3772
  "input_price_per_1m": 3,
3773
  "output_price_per_1m": 15,
 
3780
  "hf_private": true
3781
  },
3782
  {
3783
+ "name": "vertex/claude-sonnet-4",
3784
  "type": "vision",
3785
  "input_price_per_1m": 3,
3786
  "output_price_per_1m": 15,
3787
  "currency": "USD",
3788
  "capabilities": [
3789
+ "vision",
3790
+ "reasoning"
3791
  ],
3792
  "context_window": 200000,
3793
  "hf_private": true
3794
  },
3795
  {
3796
+ "name": "vertex/claude-3-7-sonnet",
3797
  "type": "vision",
3798
  "input_price_per_1m": 3,
3799
  "output_price_per_1m": 15,
 
3806
  "hf_private": true
3807
  },
3808
  {
3809
+ "name": "vertex/claude-3-5-sonnet",
3810
+ "type": "vision",
3811
  "input_price_per_1m": 3,
3812
+ "output_price_per_1m": 15,
3813
  "currency": "USD",
 
3814
  "capabilities": [
3815
+ "vision"
 
3816
  ],
3817
+ "context_window": 200000,
3818
+ "hf_private": true
 
3819
  },
3820
  {
3821
+ "name": "vertex/claude-sonnet-4-5",
3822
  "type": "vision",
3823
  "input_price_per_1m": 3,
3824
  "output_price_per_1m": 15,
3825
  "currency": "USD",
3826
  "capabilities": [
3827
+ "vision",
3828
+ "reasoning"
3829
  ],
3830
+ "context_window": 200000,
3831
  "hf_private": true
3832
  },
3833
  {
 
3872
  "hf_private": true
3873
  },
3874
  {
3875
+ "name": "xai/grok-3",
3876
+ "type": "chat",
3877
  "input_price_per_1m": 5,
3878
  "output_price_per_1m": 25,
3879
  "currency": "USD",
3880
+ "context_window": 131072,
3881
  "capabilities": [
3882
+ "tools"
 
3883
  ],
 
3884
  "hf_private": true
3885
  },
3886
  {
3887
+ "name": "bedrock/claude-opus-4-5",
3888
  "type": "vision",
3889
  "input_price_per_1m": 5,
3890
  "output_price_per_1m": 25,
 
3893
  "vision",
3894
  "reasoning"
3895
  ],
3896
+ "context_window": 200000,
3897
  "hf_private": true
3898
  },
3899
  {
3900
+ "name": "bedrock/claude-opus-4-6",
3901
  "type": "vision",
3902
  "input_price_per_1m": 5,
3903
  "output_price_per_1m": 25,
 
3906
  "vision",
3907
  "reasoning"
3908
  ],
3909
+ "context_window": 1000000,
3910
  "hf_private": true
3911
  },
3912
  {
3913
+ "name": "openai/chatgpt-4o",
3914
  "type": "vision",
3915
  "input_price_per_1m": 5,
3916
+ "output_price_per_1m": 15,
3917
  "currency": "USD",
3918
  "capabilities": [
3919
+ "vision"
 
3920
  ],
3921
+ "context_window": 128000,
3922
  "hf_private": true
3923
  },
3924
  {
3925
+ "name": "vertex/claude-opus-4-5",
3926
  "type": "vision",
3927
  "input_price_per_1m": 5,
3928
+ "output_price_per_1m": 25,
3929
  "currency": "USD",
3930
  "capabilities": [
3931
+ "vision",
3932
+ "reasoning"
3933
  ],
3934
+ "context_window": 200000,
3935
  "hf_private": true
3936
  },
3937
  {
3938
+ "name": "vertex/claude-opus-4-6",
3939
+ "type": "vision",
3940
  "input_price_per_1m": 5,
3941
  "output_price_per_1m": 25,
3942
  "currency": "USD",
 
3943
  "capabilities": [
3944
+ "vision",
3945
+ "reasoning"
3946
  ],
3947
+ "context_window": 1000000,
3948
  "hf_private": true
3949
  },
3950
  {
 
3961
  "hf_private": true
3962
  },
3963
  {
3964
+ "name": "anthropic/claude-opus-4-1",
3965
  "type": "vision",
3966
  "input_price_per_1m": 15,
3967
  "output_price_per_1m": 75,
 
3974
  "hf_private": true
3975
  },
3976
  {
3977
+ "name": "anthropic/claude-opus-4",
3978
  "type": "vision",
3979
  "input_price_per_1m": 15,
3980
  "output_price_per_1m": 75,
 
3987
  "hf_private": true
3988
  },
3989
  {
3990
+ "name": "bedrock/claude-opus-4",
3991
  "type": "vision",
3992
  "input_price_per_1m": 15,
3993
  "output_price_per_1m": 75,
 
4013
  "hf_private": true
4014
  },
4015
  {
4016
+ "name": "openai/o1",
4017
  "type": "vision",
4018
  "input_price_per_1m": 15,
4019
+ "output_price_per_1m": 60,
4020
  "currency": "USD",
4021
  "capabilities": [
4022
  "vision",
 
4026
  "hf_private": true
4027
  },
4028
  {
4029
+ "name": "coding/claude-opus-4-20250514",
4030
  "type": "vision",
4031
  "input_price_per_1m": 15,
4032
  "output_price_per_1m": 75,
 
4039
  "hf_private": true
4040
  },
4041
  {
4042
+ "name": "vertex/claude-opus-4",
4043
  "type": "vision",
4044
  "input_price_per_1m": 15,
4045
  "output_price_per_1m": 75,
 
4052
  "hf_private": true
4053
  },
4054
  {
4055
+ "name": "vertex/claude-opus-4-1",
4056
  "type": "vision",
4057
  "input_price_per_1m": 15,
4058
+ "output_price_per_1m": 75,
4059
  "currency": "USD",
4060
  "capabilities": [
4061
  "vision",
 
4091
  "hf_private": true
4092
  },
4093
  {
4094
+ "name": "openai/gpt-5.4-pro",
4095
  "type": "vision",
4096
  "input_price_per_1m": 30,
4097
  "output_price_per_1m": 180,
 
4104
  "hf_private": true
4105
  },
4106
  {
4107
+ "name": "azure/openai-responses/gpt-5.4-pro",
4108
  "type": "vision",
4109
  "input_price_per_1m": 30,
4110
  "output_price_per_1m": 180,
scripts/fetch-providers.js CHANGED
@@ -142,7 +142,9 @@ function estimateParams(config) {
142
  const l = config.num_hidden_layers || config.n_layer;
143
  const v = config.vocab_size;
144
  const i = config.intermediate_size || config.d_ff;
145
- const numExperts = config.num_local_experts || config.n_experts || 1;
 
 
146
  const modelType = (config.model_type || '').toLowerCase();
147
 
148
  if (h && l && v) {
@@ -157,9 +159,8 @@ function estimateParams(config) {
157
  // Layer parameters (Attention + MLP)
158
  const attentionParams = 4 * (h * h);
159
 
160
- // Modern architectures (Llama, Mistral, Qwen, Phi-3, Gemma) use Gated Linear Units (GLU)
161
- // which have 3 projection matrices in the MLP instead of 2.
162
- const hasGlu = ['llama', 'mistral', 'phi3', 'qwen2', 'gemma', 'gemma2'].includes(modelType);
163
  const mlpParams = (hasGlu ? 3 : 2) * h * intermediate * numExperts;
164
 
165
  const params = embedParams + l * (attentionParams + mlpParams);
@@ -173,33 +174,47 @@ async function fetchHFSize(hfId) {
173
  if (!hfId || hfId.includes(' ') || !hfId.includes('/')) return { error: 'Invalid ID' };
174
  const token = process.env.HF_TOKEN;
175
  const headers = token ? { Authorization: `Bearer ${token}` } : {};
 
 
176
  try {
177
- // 1. Get top-level metadata
178
- const data = await getJson(`https://huggingface.co/api/models/${hfId}`, { headers, retries: 1 });
179
-
180
- let params = data.safetensors?.total || data.config?.total_parameters || data.config?.model_type_params;
181
  let source = 'hf-total';
182
- if (!params && data.cardData?.model_details?.parameters) {
183
- const match = data.cardData.model_details.parameters.match(/([\d.]+)\s*[Bb]/);
184
- if (match) { params = parseFloat(match[1]) * 1_000_000_000; source = 'hf-card'; }
 
 
 
 
 
 
 
 
 
 
 
185
  }
186
 
187
  // 2. Fallback: Fetch the raw config.json file for estimation
188
- let config = data.config;
189
- if (!params && (!config || !config.hidden_size)) {
190
- try { config = await getJson(`https://huggingface.co/${hfId}/raw/main/config.json`, { headers, retries: 1 }); } catch (e) {}
 
 
 
 
 
191
  }
192
- if (!params && config) { params = estimateParams(config); source = 'hf-config-estimate'; }
193
 
194
- if (!params) return { error: 'No parameter data' };
 
195
 
196
  const b = params / 1_000_000_000;
197
  // Keep 2 decimals for small models (<1B), 1 decimal for others
198
  const size = b < 1 ? Math.round(b * 100) / 100 : Math.round(b * 10) / 10;
199
  return { size, source };
200
  } catch (e) {
201
- const isPrivate = e.message.includes('401') || e.message.includes('404');
202
- return { error: e.message, private: isPrivate };
203
  }
204
  }
205
 
@@ -222,6 +237,7 @@ async function fetchOllamaMetadata(ollamaId) {
222
 
223
  const EMBEDDER_KEYWORDS = ['embed', 'bge', 'gte', 'e5', 'stella', 'minilm', 'multilingual-mpnet'];
224
 
 
225
  const MANUAL_HF_ID_MAP = {
226
  'all minilm l12 v2': 'sentence-transformers/all-MiniLM-L12-v2',
227
  'whisper v3': 'openai/whisper-large-v3',
@@ -231,13 +247,13 @@ const MANUAL_HF_ID_MAP = {
231
  'step 3 5 flash': 'stepfun-ai/Step-3.5-Flash',
232
  'bge m3': 'BAAI/bge-m3',
233
  'bge en icl': 'BAAI/bge-en-icl',
234
- 'bge large en v1 5': 'BAAI/bge-large-en-v1.5',
235
- 'bge multilingual gemma2': 'BAAI/bge-multilingual-gemma2',
236
  'lightonocr 2': 'lightonai/LightOnOCR-2-1B',
237
  'sdxl': 'stabilityai/stable-diffusion-xl-base-1.0',
238
  'flux 1 schnell': 'black-forest-labs/FLUX.1-schnell',
239
  'flux schnell': 'black-forest-labs/FLUX.1-schnell',
240
  'paraphrase multilingual mpnet base v2': 'sentence-transformers/paraphrase-multilingual-mpnet-base-v2',
 
 
241
  'photomaker v2': 'TencentARC/PhotoMaker-V2',
242
  'canopy labs orpheus english': 'canopy-labs/orpheus-medium',
243
  'canopy labs orpheus arabic saudi': 'canopy-labs/orpheus-medium',
@@ -281,8 +297,10 @@ const MANUAL_HF_ID_MAP = {
281
  'grok 2 1212': 'xai-org/grok-2',
282
  'glm 4 6v': 'THUDM/glm-4v-9b',
283
  'glm 5 turbo': 'THUDM/glm-5-turbo',
284
- 'minimax m2 7': 'MiniMax/MiniMax-M2.7',
285
- 'minimax 01': 'MiniMax/MiniMax-Text-01',
 
 
286
  'phi 4': 'microsoft/phi-4',
287
  'flux 1 dev': 'black-forest-labs/FLUX.1-dev',
288
  'flux dev': 'black-forest-labs/FLUX.1-dev',
@@ -331,36 +349,6 @@ const MANUAL_OLLAMA_ID_MAP = {
331
  'mixtral 8x22b': 'mixtral-8x22b',
332
  };
333
 
334
- const MANUAL_SIZE_MAP = {
335
- 'BAAI/bge-m3': 0.57,
336
- 'black-forest-labs/FLUX.1-schnell': 12,
337
- 'black-forest-labs/FLUX.1-dev': 12,
338
- 'black-forest-labs/FLUX.1-pro': 12,
339
- 'black-forest-labs/FLUX.2-dev': 32,
340
- 'black-forest-labs/FLUX.2-pro': 32,
341
- 'black-forest-labs/FLUX.2-flex': 32,
342
- 'black-forest-labs/FLUX.2-max': 32,
343
- 'black-forest-labs/FLUX.2-klein-4B': 4,
344
- 'black-forest-labs/FLUX.2-klein-9B': 9,
345
- 'mistralai/Mistral-Large-Instruct-2407': 123,
346
- 'mistralai/Mistral-Large-Instruct-2411': 675,
347
- 'Alibaba/Qwen-Turbo': 14,
348
- 'Qwen/Qwen2.5-Coder-7B-Instruct': 7,
349
- 'Qwen/Qwen2.5-Coder-32B-Instruct': 32,
350
- 'Qwen/Qwen2.5-7B-Instruct': 7,
351
- 'Qwen/Qwen2-VL-7B-Instruct': 7,
352
- 'Qwen/Qwen2-VL-72B-Instruct': 72,
353
- 'deepseek-ai/DeepSeek-V3': 671,
354
- 'deepseek-ai/DeepSeek-R1': 671,
355
- 'microsoft/phi-4': 14,
356
- 'MiniMax/MiniMax-M2.7': 230,
357
- // Final public models
358
- 'TencentARC/PhotoMaker-V2': 3.1,
359
- 'stabilityai/stable-diffusion-xl-base-1.0': 2.6,
360
- 'zai-org/GLM-4.6V': 9,
361
- 'ai21labs/AI21-Jamba-Large-1.7': 52,
362
- };
363
-
364
  const PROPRIETARY_KEYWORDS = [
365
  'gpt-4', 'gpt-5', 'sonnet', 'opus', 'haiku', 'gemini', 'o1-', 'o3-', 'o4-', 'claude',
366
  'magistral', 'voxtral', 'moderation', 'embed'
@@ -391,11 +379,8 @@ async function propagateExtraData(data) {
391
  if (n === key || n.endsWith(' ' + key) || n.endsWith('/' + key)) { model.ollama_id = val; break; }
392
  }
393
  }
394
- // High-confidence size from manual map
395
- if (model.hf_id && MANUAL_SIZE_MAP[model.hf_id]) {
396
- model.size_b = MANUAL_SIZE_MAP[model.hf_id];
397
- model.size_source = 'manual';
398
- } else if (model.hf_id && !model.size_b) {
399
  const size = hfIdToSize.get(model.hf_id.toLowerCase());
400
  if (size) { model.size_b = size; model.size_source = 'benchmark'; }
401
  }
 
142
  const l = config.num_hidden_layers || config.n_layer;
143
  const v = config.vocab_size;
144
  const i = config.intermediate_size || config.d_ff;
145
+
146
+ // MoE support
147
+ const numExperts = config.num_local_experts || config.n_experts || config.num_experts || 1;
148
  const modelType = (config.model_type || '').toLowerCase();
149
 
150
  if (h && l && v) {
 
159
  // Layer parameters (Attention + MLP)
160
  const attentionParams = 4 * (h * h);
161
 
162
+ // Modern architectures (Llama, Mistral, Qwen, Phi-3, Gemma, MiniMax) use Gated Linear Units (GLU)
163
+ const hasGlu = ['llama', 'mistral', 'phi3', 'qwen2', 'gemma', 'gemma2', 'minimax'].includes(modelType);
 
164
  const mlpParams = (hasGlu ? 3 : 2) * h * intermediate * numExperts;
165
 
166
  const params = embedParams + l * (attentionParams + mlpParams);
 
174
  if (!hfId || hfId.includes(' ') || !hfId.includes('/')) return { error: 'Invalid ID' };
175
  const token = process.env.HF_TOKEN;
176
  const headers = token ? { Authorization: `Bearer ${token}` } : {};
177
+ let isPrivate = false;
178
+
179
  try {
180
+ let params = null;
 
 
 
181
  let source = 'hf-total';
182
+ let data = {};
183
+
184
+ // 1. Get top-level metadata
185
+ try {
186
+ data = await getJson(`https://huggingface.co/api/models/${hfId}`, { headers, retries: 1 });
187
+ params = data.safetensors?.total || data.config?.total_parameters || data.config?.model_type_params;
188
+
189
+ // Fallback: cardData
190
+ if (!params && data.cardData?.model_details?.parameters) {
191
+ const match = data.cardData.model_details.parameters.match(/([\d.]+)\s*[Bb]/);
192
+ if (match) { params = parseFloat(match[1]) * 1_000_000_000; source = 'hf-card'; }
193
+ }
194
+ } catch (e) {
195
+ if (e.message.includes('401') || e.message.includes('404')) isPrivate = true;
196
  }
197
 
198
  // 2. Fallback: Fetch the raw config.json file for estimation
199
+ if (!params && !isPrivate) {
200
+ try {
201
+ const config = await getJson(`https://huggingface.co/${hfId}/raw/main/config.json`, { headers, retries: 1 });
202
+ params = config.total_parameters || estimateParams(config);
203
+ source = config.total_parameters ? 'hf-total' : 'hf-config-estimate';
204
+ } catch (e) {
205
+ if (e.message.includes('401') || e.message.includes('404')) isPrivate = true;
206
+ }
207
  }
 
208
 
209
+ if (isPrivate) return { error: 'Private or Missing', private: true };
210
+ if (!params) return { error: 'No parameter data found' };
211
 
212
  const b = params / 1_000_000_000;
213
  // Keep 2 decimals for small models (<1B), 1 decimal for others
214
  const size = b < 1 ? Math.round(b * 100) / 100 : Math.round(b * 10) / 10;
215
  return { size, source };
216
  } catch (e) {
217
+ return { error: e.message };
 
218
  }
219
  }
220
 
 
237
 
238
  const EMBEDDER_KEYWORDS = ['embed', 'bge', 'gte', 'e5', 'stella', 'minilm', 'multilingual-mpnet'];
239
 
240
+ // Manual mappings for models with non-standard naming.
241
  const MANUAL_HF_ID_MAP = {
242
  'all minilm l12 v2': 'sentence-transformers/all-MiniLM-L12-v2',
243
  'whisper v3': 'openai/whisper-large-v3',
 
247
  'step 3 5 flash': 'stepfun-ai/Step-3.5-Flash',
248
  'bge m3': 'BAAI/bge-m3',
249
  'bge en icl': 'BAAI/bge-en-icl',
 
 
250
  'lightonocr 2': 'lightonai/LightOnOCR-2-1B',
251
  'sdxl': 'stabilityai/stable-diffusion-xl-base-1.0',
252
  'flux 1 schnell': 'black-forest-labs/FLUX.1-schnell',
253
  'flux schnell': 'black-forest-labs/FLUX.1-schnell',
254
  'paraphrase multilingual mpnet base v2': 'sentence-transformers/paraphrase-multilingual-mpnet-base-v2',
255
+ 'bge large en v1 5': 'BAAI/bge-large-en-v1.5',
256
+ 'bge multilingual gemma2': 'BAAI/bge-multilingual-gemma2',
257
  'photomaker v2': 'TencentARC/PhotoMaker-V2',
258
  'canopy labs orpheus english': 'canopy-labs/orpheus-medium',
259
  'canopy labs orpheus arabic saudi': 'canopy-labs/orpheus-medium',
 
297
  'grok 2 1212': 'xai-org/grok-2',
298
  'glm 4 6v': 'THUDM/glm-4v-9b',
299
  'glm 5 turbo': 'THUDM/glm-5-turbo',
300
+ 'minimax m2 7': 'MiniMaxAI/MiniMax-M2.7',
301
+ 'minimax m2 7 highspeed': 'MiniMaxAI/MiniMax-M2.7',
302
+ 'minimax 01': 'MiniMaxAI/MiniMax-Text-01',
303
+ 'minimax m2 her': 'MiniMaxAI/MiniMax-M2.7',
304
  'phi 4': 'microsoft/phi-4',
305
  'flux 1 dev': 'black-forest-labs/FLUX.1-dev',
306
  'flux dev': 'black-forest-labs/FLUX.1-dev',
 
349
  'mixtral 8x22b': 'mixtral-8x22b',
350
  };
351
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
352
  const PROPRIETARY_KEYWORDS = [
353
  'gpt-4', 'gpt-5', 'sonnet', 'opus', 'haiku', 'gemini', 'o1-', 'o3-', 'o4-', 'claude',
354
  'magistral', 'voxtral', 'moderation', 'embed'
 
379
  if (n === key || n.endsWith(' ' + key) || n.endsWith('/' + key)) { model.ollama_id = val; break; }
380
  }
381
  }
382
+
383
+ if (model.hf_id && !model.size_b) {
 
 
 
384
  const size = hfIdToSize.get(model.hf_id.toLowerCase());
385
  if (size) { model.size_b = size; model.size_source = 'benchmark'; }
386
  }