ai_usage/benchmarks.json at main · Arun-Ranjan/ai_usage · GitHub

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
{
  "code_review": {
    "task": "code review (Python file, ~60 lines)",
    "models": [
      {
        "name": "GPT-4",
        "tokens_per_review": 700,
        "cost_per_review": 0.021,
        "quality": "SOTA, highest accuracy for complex tasks",
        "source": "https://openai.com/pricing"
      },
      {
        "name": "Claude 2",
        "tokens_per_review": 700,
        "cost_per_review": 0.0168,
        "quality": "Competitive with GPT-3.5, capable on code review",
        "source": "https://www.anthropic.com/api"
      },
      {
        "name": "Gemini 1.5 Pro",
        "tokens_per_review": 700,
        "cost_per_review": 0.0035,
        "quality": "Competitive with GPT-3.5/Claude 2",
        "source": "https://cloud.google.com/vertex-ai/docs/generative-ai/pricing"
      },
      {
        "name": "Claude Instant",
        "tokens_per_review": 700,
        "cost_per_review": 0.00467,
        "quality": "Very fast, less accurate for complex review, best for routine checks",
        "source": "https://www.anthropic.com/api"
      },
      {
        "name": "CodeBERT (fine-tuned)",
        "tokens_per_review": 120,
        "cost_per_review": 0.00036,
        "quality": "Very high, within 1-2 points of GPT-4 for routine bug/style review",
        "source": "https://arxiv.org/abs/2107.03374"
      },
      {
        "name": "CodeT5+ (fine-tuned)",
        "tokens_per_review": 100,
        "cost_per_review": 0.0001,
        "quality": "High, nearly matches GPT-4 for focused code review tasks",
        "source": "https://arxiv.org/abs/2302.09413"
      }
    ],
    "sources": [
      "https://openai.com/pricing",
      "https://arxiv.org/abs/2107.03374",
      "https://arxiv.org/abs/2302.09413",
      "https://github.com/microsoft/CodeXGLUE",
      "https://www.anthropic.com/api",
      "https://cloud.google.com/vertex-ai/docs/generative-ai/pricing"
    ],
    "summary": "Gemini Pro offers competitive accuracy to Claude 2 for code reviews at a significantly lower price per run. Fine-tuned models remain most efficient for routine code checks."
  },
  "code_summarization": {
    "task": "code summarization (Python function, ~30 lines)",
    "models": [
      {
        "name": "GPT-4",
        "tokens_per_summary": 300,
        "cost_per_summary": 0.009,
        "quality": "SOTA, excellent natural language summaries",
        "source": "https://openai.com/pricing"
      },
      {
        "name": "Claude 2",
        "tokens_per_summary": 300,
        "cost_per_summary": 0.0072,
        "quality": "Competitive, strong summarization capabilities",
        "source": "https://www.anthropic.com/api"
      },
      {
        "name": "Gemini 1.5 Pro",
        "tokens_per_summary": 300,
        "cost_per_summary": 0.0015,
        "quality": "Competitive with Claude 2",
        "source": "https://cloud.google.com/vertex-ai/docs/generative-ai/pricing"
      },
      {
        "name": "Claude Instant",
        "tokens_per_summary": 300,
        "cost_per_summary": 0.00165,
        "quality": "Fast, good for basic summarization",
        "source": "https://www.anthropic.com/api"
      },
      {
        "name": "CodeBERT",
        "tokens_per_summary": 55,
        "cost_per_summary": 0.000165,
        "quality": "High, concise and accurate summaries",
        "source": "https://arxiv.org/abs/2107.03374"
      },
      {
        "name": "CodeT5 (fine-tuned)",
        "tokens_per_summary": 42,
        "cost_per_summary": 0.000042,
        "quality": "Very high, optimized for code summarization",
        "source": "https://arxiv.org/abs/2302.09413"
      }
    ],
    "sources": [
      "https://openai.com/pricing",
      "https://arxiv.org/abs/2107.03374",
      "https://arxiv.org/abs/2302.09413",
      "https://www.anthropic.com/api",
      "https://cloud.google.com/vertex-ai/docs/generative-ai/pricing"
    ],
    "summary": "Fine-tuned models like CodeT5 provide excellent summarization quality at a fraction of the cost of general LLMs. Gemini Pro offers a good balance for teams needing general-purpose capabilities."
  },
  "bug_detection": {
    "task": "bug detection (Python script, ~40 lines)",
    "models": [
      {
        "name": "GPT-4",
        "tokens_per_detection": 400,
        "cost_per_detection": 0.012,
        "quality": "SOTA, excellent at finding subtle bugs",
        "source": "https://openai.com/pricing"
      },
      {
        "name": "Claude 2",
        "tokens_per_detection": 400,
        "cost_per_detection": 0.0096,
        "quality": "Competitive, strong bug detection",
        "source": "https://www.anthropic.com/api"
      },
      {
        "name": "Gemini 1.5 Pro",
        "tokens_per_detection": 400,
        "cost_per_detection": 0.002,
        "quality": "Competitive with Claude 2",
        "source": "https://cloud.google.com/vertex-ai/docs/generative-ai/pricing"
      },
      {
        "name": "Claude Instant",
        "tokens_per_detection": 400,
        "cost_per_detection": 0.0022,
        "quality": "Good for common bug patterns",
        "source": "https://www.anthropic.com/api"
      },
      {
        "name": "CodeBERT",
        "tokens_per_detection": 70,
        "cost_per_detection": 0.00021,
        "quality": "High, effective for standard bug types",
        "source": "https://arxiv.org/abs/2107.03374"
      },
      {
        "name": "BugLab (fine-tuned)",
        "tokens_per_detection": 60,
        "cost_per_detection": 0.00006,
        "quality": "Very high, specialized for bug detection",
        "source": "https://arxiv.org/abs/2107.10864"
      }
    ],
    "sources": [
      "https://openai.com/pricing",
      "https://arxiv.org/abs/2107.03374",
      "https://arxiv.org/abs/2107.10864",
      "https://www.anthropic.com/api",
      "https://cloud.google.com/vertex-ai/docs/generative-ai/pricing"
    ],
    "summary": "BugLab and CodeBERT offer exceptional bug detection at minimal cost. For complex or novel bug patterns, GPT-4 remains the gold standard."
  },
  "code_qa": {
    "task": "code Q&A (e.g., 'What does this function do?')",
    "models": [
      {
        "name": "GPT-4",
        "tokens_per_qa": 250,
        "cost_per_qa": 0.0075,
        "quality": "SOTA, excellent explanations",
        "source": "https://openai.com/pricing"
      },
      {
        "name": "Claude 2",
        "tokens_per_qa": 250,
        "cost_per_qa": 0.006,
        "quality": "Competitive, clear explanations",
        "source": "https://www.anthropic.com/api"
      },
      {
        "name": "Gemini 1.5 Pro",
        "tokens_per_qa": 250,
        "cost_per_qa": 0.00125,
        "quality": "Competitive with Claude 2",
        "source": "https://cloud.google.com/vertex-ai/docs/generative-ai/pricing"
      },
      {
        "name": "Claude Instant",
        "tokens_per_qa": 250,
        "cost_per_qa": 0.001375,
        "quality": "Good for straightforward Q&A",
        "source": "https://www.anthropic.com/api"
      },
      {
        "name": "CodeT5+ (fine-tuned)",
        "tokens_per_qa": 40,
        "cost_per_qa": 0.00004,
        "quality": "High, concise and accurate answers",
        "source": "https://arxiv.org/abs/2302.09413"
      }
    ],
    "sources": [
      "https://openai.com/pricing",
      "https://arxiv.org/abs/2302.09413",
      "https://www.anthropic.com/api",
      "https://cloud.google.com/vertex-ai/docs/generative-ai/pricing"
    ],
    "summary": "CodeT5+ delivers high-quality answers at extremely low cost. General LLMs excel when questions require broader context or creative reasoning."
  }
}