/ mlflow / utils / model_catalog / replicate.json
replicate.json
  1  {
  2    "schema_version": "1.0",
  3    "models": {
  4      "anthropic/claude-3.5-haiku": {
  5        "mode": "chat",
  6        "pricing": {
  7          "input_per_million_tokens": 1.0,
  8          "output_per_million_tokens": 5.0
  9        },
 10        "capabilities": {
 11          "function_calling": true,
 12          "vision": true,
 13          "reasoning": false,
 14          "prompt_caching": true,
 15          "response_schema": true
 16        },
 17        "last_updated_at": "2026-04-24"
 18      },
 19      "anthropic/claude-3.5-sonnet": {
 20        "mode": "chat",
 21        "pricing": {
 22          "input_per_million_tokens": 3.75,
 23          "output_per_million_tokens": 18.75
 24        },
 25        "capabilities": {
 26          "function_calling": true,
 27          "vision": true,
 28          "reasoning": false,
 29          "prompt_caching": true,
 30          "response_schema": true
 31        },
 32        "last_updated_at": "2026-04-24"
 33      },
 34      "anthropic/claude-3.7-sonnet": {
 35        "mode": "chat",
 36        "pricing": {
 37          "input_per_million_tokens": 3.0,
 38          "output_per_million_tokens": 15.0
 39        },
 40        "capabilities": {
 41          "function_calling": true,
 42          "vision": true,
 43          "reasoning": false,
 44          "prompt_caching": true,
 45          "response_schema": true
 46        },
 47        "last_updated_at": "2026-04-24"
 48      },
 49      "anthropic/claude-4-sonnet": {
 50        "mode": "chat",
 51        "pricing": {
 52          "input_per_million_tokens": 3.0,
 53          "output_per_million_tokens": 15.0
 54        },
 55        "capabilities": {
 56          "function_calling": true,
 57          "vision": true,
 58          "reasoning": false,
 59          "prompt_caching": true,
 60          "response_schema": true
 61        },
 62        "last_updated_at": "2026-04-24"
 63      },
 64      "anthropic/claude-4.5-haiku": {
 65        "mode": "chat",
 66        "pricing": {
 67          "input_per_million_tokens": 1.0,
 68          "output_per_million_tokens": 5.0
 69        },
 70        "capabilities": {
 71          "function_calling": true,
 72          "vision": true,
 73          "reasoning": false,
 74          "prompt_caching": true,
 75          "response_schema": true
 76        },
 77        "last_updated_at": "2026-04-24"
 78      },
 79      "anthropic/claude-4.5-sonnet": {
 80        "mode": "chat",
 81        "pricing": {
 82          "input_per_million_tokens": 3.0,
 83          "output_per_million_tokens": 15.0
 84        },
 85        "capabilities": {
 86          "function_calling": true,
 87          "vision": true,
 88          "reasoning": false,
 89          "prompt_caching": true,
 90          "response_schema": true
 91        },
 92        "last_updated_at": "2026-04-24"
 93      },
 94      "deepseek-ai/deepseek-r1": {
 95        "mode": "chat",
 96        "context_window": {
 97          "max_input": 65536,
 98          "max_output": 8192,
 99          "max_tokens": 8192
100        },
101        "pricing": {
102          "input_per_million_tokens": 3.75,
103          "output_per_million_tokens": 10.0
104        },
105        "capabilities": {
106          "function_calling": false,
107          "vision": false,
108          "reasoning": true,
109          "prompt_caching": false,
110          "response_schema": false
111        },
112        "last_updated_at": "2026-04-24"
113      },
114      "deepseek-ai/deepseek-v3": {
115        "mode": "chat",
116        "context_window": {
117          "max_input": 65536,
118          "max_output": 8192,
119          "max_tokens": 8192
120        },
121        "pricing": {
122          "input_per_million_tokens": 1.45,
123          "output_per_million_tokens": 1.45
124        },
125        "capabilities": {
126          "function_calling": true,
127          "vision": false,
128          "reasoning": false,
129          "prompt_caching": false,
130          "response_schema": false
131        },
132        "last_updated_at": "2026-04-24"
133      },
134      "deepseek-ai/deepseek-v3.1": {
135        "mode": "chat",
136        "context_window": {
137          "max_input": 163840,
138          "max_output": 163840,
139          "max_tokens": 163840
140        },
141        "pricing": {
142          "input_per_million_tokens": 0.672,
143          "output_per_million_tokens": 2.016
144        },
145        "capabilities": {
146          "function_calling": true,
147          "vision": false,
148          "reasoning": true,
149          "prompt_caching": false,
150          "response_schema": false
151        },
152        "last_updated_at": "2026-04-24"
153      },
154      "google/gemini-2.5-flash": {
155        "mode": "chat",
156        "pricing": {
157          "input_per_million_tokens": 2.5,
158          "output_per_million_tokens": 2.5
159        },
160        "capabilities": {
161          "function_calling": true,
162          "vision": true,
163          "reasoning": false,
164          "prompt_caching": false,
165          "response_schema": true
166        },
167        "last_updated_at": "2026-04-24"
168      },
169      "google/gemini-3-pro": {
170        "mode": "chat",
171        "pricing": {
172          "input_per_million_tokens": 2.0,
173          "output_per_million_tokens": 12.0
174        },
175        "capabilities": {
176          "function_calling": true,
177          "vision": true,
178          "reasoning": false,
179          "prompt_caching": false,
180          "response_schema": true
181        },
182        "last_updated_at": "2026-04-24"
183      },
184      "gpt-oss-20b": {
185        "mode": "chat",
186        "pricing": {
187          "input_per_million_tokens": 0.09,
188          "output_per_million_tokens": 0.36
189        },
190        "capabilities": {
191          "function_calling": true,
192          "vision": false,
193          "reasoning": false,
194          "prompt_caching": false,
195          "response_schema": false
196        },
197        "last_updated_at": "2026-04-24"
198      },
199      "ibm-granite/granite-3.3-8b-instruct": {
200        "mode": "chat",
201        "pricing": {
202          "input_per_million_tokens": 0.03,
203          "output_per_million_tokens": 0.25
204        },
205        "capabilities": {
206          "function_calling": true,
207          "vision": false,
208          "reasoning": false,
209          "prompt_caching": false,
210          "response_schema": false
211        },
212        "last_updated_at": "2026-04-24"
213      },
214      "meta/llama-2-13b": {
215        "mode": "chat",
216        "context_window": {
217          "max_input": 4096,
218          "max_output": 4096,
219          "max_tokens": 4096
220        },
221        "pricing": {
222          "input_per_million_tokens": 0.1,
223          "output_per_million_tokens": 0.5
224        },
225        "capabilities": {
226          "function_calling": false,
227          "vision": false,
228          "reasoning": false,
229          "prompt_caching": false,
230          "response_schema": false
231        },
232        "last_updated_at": "2026-04-24"
233      },
234      "meta/llama-2-13b-chat": {
235        "mode": "chat",
236        "context_window": {
237          "max_input": 4096,
238          "max_output": 4096,
239          "max_tokens": 4096
240        },
241        "pricing": {
242          "input_per_million_tokens": 0.1,
243          "output_per_million_tokens": 0.5
244        },
245        "capabilities": {
246          "function_calling": false,
247          "vision": false,
248          "reasoning": false,
249          "prompt_caching": false,
250          "response_schema": false
251        },
252        "last_updated_at": "2026-04-24"
253      },
254      "meta/llama-2-70b": {
255        "mode": "chat",
256        "context_window": {
257          "max_input": 4096,
258          "max_output": 4096,
259          "max_tokens": 4096
260        },
261        "pricing": {
262          "input_per_million_tokens": 0.65,
263          "output_per_million_tokens": 2.75
264        },
265        "capabilities": {
266          "function_calling": false,
267          "vision": false,
268          "reasoning": false,
269          "prompt_caching": false,
270          "response_schema": false
271        },
272        "last_updated_at": "2026-04-24"
273      },
274      "meta/llama-2-70b-chat": {
275        "mode": "chat",
276        "context_window": {
277          "max_input": 4096,
278          "max_output": 4096,
279          "max_tokens": 4096
280        },
281        "pricing": {
282          "input_per_million_tokens": 0.65,
283          "output_per_million_tokens": 2.75
284        },
285        "capabilities": {
286          "function_calling": false,
287          "vision": false,
288          "reasoning": false,
289          "prompt_caching": false,
290          "response_schema": false
291        },
292        "last_updated_at": "2026-04-24"
293      },
294      "meta/llama-2-7b": {
295        "mode": "chat",
296        "context_window": {
297          "max_input": 4096,
298          "max_output": 4096,
299          "max_tokens": 4096
300        },
301        "pricing": {
302          "input_per_million_tokens": 0.05,
303          "output_per_million_tokens": 0.25
304        },
305        "capabilities": {
306          "function_calling": false,
307          "vision": false,
308          "reasoning": false,
309          "prompt_caching": false,
310          "response_schema": false
311        },
312        "last_updated_at": "2026-04-24"
313      },
314      "meta/llama-2-7b-chat": {
315        "mode": "chat",
316        "context_window": {
317          "max_input": 4096,
318          "max_output": 4096,
319          "max_tokens": 4096
320        },
321        "pricing": {
322          "input_per_million_tokens": 0.05,
323          "output_per_million_tokens": 0.25
324        },
325        "capabilities": {
326          "function_calling": false,
327          "vision": false,
328          "reasoning": false,
329          "prompt_caching": false,
330          "response_schema": false
331        },
332        "last_updated_at": "2026-04-24"
333      },
334      "meta/llama-3-70b": {
335        "mode": "chat",
336        "context_window": {
337          "max_input": 8192,
338          "max_output": 8192,
339          "max_tokens": 8192
340        },
341        "pricing": {
342          "input_per_million_tokens": 0.65,
343          "output_per_million_tokens": 2.75
344        },
345        "capabilities": {
346          "function_calling": false,
347          "vision": false,
348          "reasoning": false,
349          "prompt_caching": false,
350          "response_schema": false
351        },
352        "last_updated_at": "2026-04-24"
353      },
354      "meta/llama-3-70b-instruct": {
355        "mode": "chat",
356        "context_window": {
357          "max_input": 8192,
358          "max_output": 8192,
359          "max_tokens": 8192
360        },
361        "pricing": {
362          "input_per_million_tokens": 0.65,
363          "output_per_million_tokens": 2.75
364        },
365        "capabilities": {
366          "function_calling": false,
367          "vision": false,
368          "reasoning": false,
369          "prompt_caching": false,
370          "response_schema": false
371        },
372        "last_updated_at": "2026-04-24"
373      },
374      "meta/llama-3-8b": {
375        "mode": "chat",
376        "context_window": {
377          "max_input": 8086,
378          "max_output": 8086,
379          "max_tokens": 8086
380        },
381        "pricing": {
382          "input_per_million_tokens": 0.05,
383          "output_per_million_tokens": 0.25
384        },
385        "capabilities": {
386          "function_calling": false,
387          "vision": false,
388          "reasoning": false,
389          "prompt_caching": false,
390          "response_schema": false
391        },
392        "last_updated_at": "2026-04-24"
393      },
394      "meta/llama-3-8b-instruct": {
395        "mode": "chat",
396        "context_window": {
397          "max_input": 8086,
398          "max_output": 8086,
399          "max_tokens": 8086
400        },
401        "pricing": {
402          "input_per_million_tokens": 0.05,
403          "output_per_million_tokens": 0.25
404        },
405        "capabilities": {
406          "function_calling": false,
407          "vision": false,
408          "reasoning": false,
409          "prompt_caching": false,
410          "response_schema": false
411        },
412        "last_updated_at": "2026-04-24"
413      },
414      "mistralai/mistral-7b-instruct-v0.2": {
415        "mode": "chat",
416        "context_window": {
417          "max_input": 4096,
418          "max_output": 4096,
419          "max_tokens": 4096
420        },
421        "pricing": {
422          "input_per_million_tokens": 0.05,
423          "output_per_million_tokens": 0.25
424        },
425        "capabilities": {
426          "function_calling": false,
427          "vision": false,
428          "reasoning": false,
429          "prompt_caching": false,
430          "response_schema": false
431        },
432        "last_updated_at": "2026-04-24"
433      },
434      "mistralai/mistral-7b-v0.1": {
435        "mode": "chat",
436        "context_window": {
437          "max_input": 4096,
438          "max_output": 4096,
439          "max_tokens": 4096
440        },
441        "pricing": {
442          "input_per_million_tokens": 0.05,
443          "output_per_million_tokens": 0.25
444        },
445        "capabilities": {
446          "function_calling": false,
447          "vision": false,
448          "reasoning": false,
449          "prompt_caching": false,
450          "response_schema": false
451        },
452        "last_updated_at": "2026-04-24"
453      },
454      "mistralai/mixtral-8x7b-instruct-v0.1": {
455        "mode": "chat",
456        "context_window": {
457          "max_input": 4096,
458          "max_output": 4096,
459          "max_tokens": 4096
460        },
461        "pricing": {
462          "input_per_million_tokens": 0.3,
463          "output_per_million_tokens": 1.0
464        },
465        "capabilities": {
466          "function_calling": false,
467          "vision": false,
468          "reasoning": false,
469          "prompt_caching": false,
470          "response_schema": false
471        },
472        "last_updated_at": "2026-04-24"
473      },
474      "openai/gpt-4.1": {
475        "mode": "chat",
476        "pricing": {
477          "input_per_million_tokens": 2.0,
478          "output_per_million_tokens": 8.0
479        },
480        "capabilities": {
481          "function_calling": true,
482          "vision": true,
483          "reasoning": false,
484          "prompt_caching": false,
485          "response_schema": true
486        },
487        "last_updated_at": "2026-04-24"
488      },
489      "openai/gpt-4.1-mini": {
490        "mode": "chat",
491        "pricing": {
492          "input_per_million_tokens": 0.4,
493          "output_per_million_tokens": 1.6
494        },
495        "capabilities": {
496          "function_calling": true,
497          "vision": true,
498          "reasoning": false,
499          "prompt_caching": false,
500          "response_schema": true
501        },
502        "last_updated_at": "2026-04-24"
503      },
504      "openai/gpt-4.1-nano": {
505        "mode": "chat",
506        "pricing": {
507          "input_per_million_tokens": 0.1,
508          "output_per_million_tokens": 0.4
509        },
510        "capabilities": {
511          "function_calling": true,
512          "vision": false,
513          "reasoning": false,
514          "prompt_caching": false,
515          "response_schema": false
516        },
517        "last_updated_at": "2026-04-24"
518      },
519      "openai/gpt-4o": {
520        "mode": "chat",
521        "pricing": {
522          "input_per_million_tokens": 2.5,
523          "output_per_million_tokens": 10.0
524        },
525        "capabilities": {
526          "function_calling": true,
527          "vision": true,
528          "reasoning": false,
529          "prompt_caching": false,
530          "response_schema": true
531        },
532        "last_updated_at": "2026-04-24"
533      },
534      "openai/gpt-4o-mini": {
535        "mode": "chat",
536        "pricing": {
537          "input_per_million_tokens": 0.15,
538          "output_per_million_tokens": 0.6
539        },
540        "capabilities": {
541          "function_calling": true,
542          "vision": true,
543          "reasoning": false,
544          "prompt_caching": false,
545          "response_schema": true
546        },
547        "last_updated_at": "2026-04-24"
548      },
549      "openai/gpt-5": {
550        "mode": "chat",
551        "pricing": {
552          "input_per_million_tokens": 1.25,
553          "output_per_million_tokens": 10.0
554        },
555        "capabilities": {
556          "function_calling": true,
557          "vision": true,
558          "reasoning": false,
559          "prompt_caching": false,
560          "response_schema": true
561        },
562        "last_updated_at": "2026-04-24"
563      },
564      "openai/gpt-5-mini": {
565        "mode": "chat",
566        "pricing": {
567          "input_per_million_tokens": 0.25,
568          "output_per_million_tokens": 2.0
569        },
570        "capabilities": {
571          "function_calling": true,
572          "vision": true,
573          "reasoning": false,
574          "prompt_caching": false,
575          "response_schema": true
576        },
577        "last_updated_at": "2026-04-24"
578      },
579      "openai/gpt-5-nano": {
580        "mode": "chat",
581        "pricing": {
582          "input_per_million_tokens": 0.05,
583          "output_per_million_tokens": 0.4
584        },
585        "capabilities": {
586          "function_calling": true,
587          "vision": false,
588          "reasoning": false,
589          "prompt_caching": false,
590          "response_schema": false
591        },
592        "last_updated_at": "2026-04-24"
593      },
594      "openai/gpt-oss-120b": {
595        "mode": "chat",
596        "pricing": {
597          "input_per_million_tokens": 0.18,
598          "output_per_million_tokens": 0.72
599        },
600        "capabilities": {
601          "function_calling": true,
602          "vision": false,
603          "reasoning": false,
604          "prompt_caching": false,
605          "response_schema": false
606        },
607        "last_updated_at": "2026-04-24"
608      },
609      "openai/o1": {
610        "mode": "chat",
611        "pricing": {
612          "input_per_million_tokens": 15.0,
613          "output_per_million_tokens": 60.0
614        },
615        "capabilities": {
616          "function_calling": false,
617          "vision": false,
618          "reasoning": true,
619          "prompt_caching": false,
620          "response_schema": false
621        },
622        "last_updated_at": "2026-04-24"
623      },
624      "openai/o1-mini": {
625        "mode": "chat",
626        "pricing": {
627          "input_per_million_tokens": 1.1,
628          "output_per_million_tokens": 4.4
629        },
630        "capabilities": {
631          "function_calling": false,
632          "vision": false,
633          "reasoning": true,
634          "prompt_caching": false,
635          "response_schema": false
636        },
637        "last_updated_at": "2026-04-24"
638      },
639      "openai/o4-mini": {
640        "mode": "chat",
641        "pricing": {
642          "input_per_million_tokens": 1.0,
643          "output_per_million_tokens": 4.0
644        },
645        "capabilities": {
646          "function_calling": false,
647          "vision": false,
648          "reasoning": true,
649          "prompt_caching": false,
650          "response_schema": false
651        },
652        "last_updated_at": "2026-04-24"
653      },
654      "qwen/qwen3-235b-a22b-instruct-2507": {
655        "mode": "chat",
656        "pricing": {
657          "input_per_million_tokens": 0.264,
658          "output_per_million_tokens": 1.06
659        },
660        "capabilities": {
661          "function_calling": true,
662          "vision": false,
663          "reasoning": false,
664          "prompt_caching": false,
665          "response_schema": false
666        },
667        "last_updated_at": "2026-04-24"
668      },
669      "xai/grok-4": {
670        "mode": "chat",
671        "pricing": {
672          "input_per_million_tokens": 7.2,
673          "output_per_million_tokens": 36.0
674        },
675        "capabilities": {
676          "function_calling": true,
677          "vision": false,
678          "reasoning": false,
679          "prompt_caching": false,
680          "response_schema": false
681        },
682        "last_updated_at": "2026-04-24"
683      }
684    }
685  }