@@ -80,6 +80,7 @@ def parse_args():
8080 "Llama 4 Maverick 17B" : "closed" ,
8181 "Claude 4 Opus" : "closed" ,
8282 "OpenAI O3" : "closed" ,
83+ "OpenAI o3-pro" : "closed" ,
8384}
8485
8586models_in_playground = set ([
@@ -96,6 +97,7 @@ def parse_args():
9697 "GPT-4.1 Nano" ,
9798 "ChatGPT-4o" ,
9899 "OpenAI O4 Mini" ,
100+ "OpenAI o3-pro"
99101])
100102
101103
@@ -124,7 +126,6 @@ def slugify(value):
124126 return value
125127
126128 logos = {
127- "DeepSeek R1" : "https://encrypted-tbn0.gstatic.com/images?q=tbn:ANd9GcREUufvK7RFFGsbQTZQq9BZgFASSDa7xOTb1A&s" ,
128129 "Llama 3 11B Vision" : "https://signsalad.com/wp-content/uploads/2021/11/Screenshot-2021-11-03-at-12.14.11.png" ,
129130 "Llama 3 70B" : "https://signsalad.com/wp-content/uploads/2021/11/Screenshot-2021-11-03-at-12.14.11.png" ,
130131 "Llama 4 Maverick 17B" : "https://signsalad.com/wp-content/uploads/2021/11/Screenshot-2021-11-03-at-12.14.11.png" ,
@@ -158,6 +159,7 @@ def slugify(value):
158159 "Gemma 3 4B" : "https://www.google.com/favicon.ico" ,
159160 "Gemini 1.5 Flash" : "https://www.google.com/favicon.ico" ,
160161 "Gemini 1.5 Pro" : "https://www.google.com/favicon.ico" ,
162+ "OpenAI o3-pro" : "https://openai.com/favicon.ico" ,
161163 "Arcee.ai Spotlight" : "https://cdn.prod.website-files.com/6781a10424493fe352bc6cb5/678e92cb5d392e76c953e690_Favicon.png" ,
162164 "Phi 4 Multimodal" : "https://upload.wikimedia.org/wikipedia/commons/thumb/4/44/Microsoft_logo.svg/1024px-Microsoft_logo.svg.png?20210729021049" ,
163165 }
@@ -260,6 +262,7 @@ def run_model_with_prompt(model_name, model, assessment):
260262 "Gemini 1.5 Flash" : "" ,
261263 "Gemini 1.5 Pro" : "" ,
262264 "Arcee.ai Spotlight" : "" ,
265+ "OpenAI o3-pro" : "" ,
263266 }
264267 # load from saved_results
265268 assessments_by_model = final_results ["assessments_by_model" ]
@@ -294,6 +297,7 @@ def run_model_with_prompt(model_name, model, assessment):
294297 base_url = "https://router.huggingface.co/together/v1" ,
295298 api_key = os .environ .get ("HUGGINGFACE_API_KEY" ),
296299 ),
300+ "OpenAI o3-pro" : OpenAIModel (model_id = "o3-pro" ),
297301 "Claude 3.7 Sonnet" : AnthropicModel (model_id = "claude-3-7-sonnet-20250219" ),
298302 "Claude 3.5 Haiku" : AnthropicModel (model_id = "claude-3-5-haiku-20241022" ),
299303 "Gemini 2.5 Pro Preview" : GeminiModel (model_id = "gemini-2.5-pro-preview-03-25" ),
@@ -304,11 +308,6 @@ def run_model_with_prompt(model_name, model, assessment):
304308 "Gemini 2.5 Flash Preview" : GeminiModel (model_id = "gemini-2.5-flash-preview-04-17" ),
305309 "Cohere Aya Vision 8B" : CohereModel (model_id = "c4ai-aya-vision-8b" ),
306310 "Cohere Aya Vision 32B" : CohereModel (model_id = "c4ai-aya-vision-32b" ),
307- "DeepSeek R1" : CustomOpenAIModel (
308- model_id = "deepseek-ai/DeepSeek-R1" ,
309- base_url = "https://router.huggingface.co/hyperbolic/v1" ,
310- api_key = os .environ .get ("HUGGINGFACE_API_KEY" ),
311- ),
312311 "Qwen 2.5 VL 7B" : CustomOpenAIModel (
313312 model_id = "Qwen/Qwen2.5-VL-7B-Instruct" ,
314313 base_url = "https://router.huggingface.co/hyperbolic/v1" ,
0 commit comments