o3 and o3-mini

[simonw]

Released today:

• https://platform.openai.com/docs/models/o4-mini
• https://platform.openai.com/docs/models/o3

[simonw]

I could derive the full model date-based IDs like this:

% llm -m openai/o3 hi
Error: Error code: 400 - {'error': {'message': 'Your organization must be verified to stream this model. Please go to: https://platform.openai.com/settings/organization/general and click on Verify Organization.', 'type': 'invalid_request_error', 'param': 'stream', 'code': 'unsupported_value'}}
% llm -m openai/o3 hi --no-stream
Error: Error code: 404 - {'error': {'message': 'The model `o3-2025-04-16` does not exist or you do not have access to it.', 'type': 'invalid_request_error', 'param': None, 'code': 'model_not_found'}}
% llm -m openai/o4-mini hi       
Error: The model `o4-mini-2025-04-16` does not exist or you do not have access to it.
% llm -m openai/o3 hi --no-stream
Error: Error code: 404 - {'error': {'message': 'The model `o3-2025-04-16` does not exist or you do not have access to it.', 'type': 'invalid_request_error', 'param': None, 'code': 'model_not_found'}}

[simonw]

Here's what I have so far:

diff --git a/llm_openai.py b/llm_openai.py
index d6d419b..ac5f951 100644
--- a/llm_openai.py
+++ b/llm_openai.py
@@ -32,6 +32,11 @@ def register_models(register):
         "gpt-4.1-mini-2025-04-14": {"vision": True},
         "gpt-4.1-nano": {"vision": True},
         "gpt-4.1-nano-2025-04-14": {"vision": True},
+        # April 16th 2025
+        "o3": {"vision": True, "reasoning": True},
+        "o3-2025-04-16": {"vision": True, "reasoning": True},
+        "o4-mini": {"vision": True, "reasoning": True},
+        "o4-mini-2025-04-16": {"vision": True, "reasoning": True},
     }
     for model_id, options in models.items():
         register(

That "Your organization must be verified to stream this model" message suggests I should set mark o3 as not supporting streaming, and maybe register an o3-streaming separate model ID that people with verified organizations can use.

[simonw]

Looks like the models are not live yet, when they are I'll test them like this:

llm -m openai/o3 "say hi in five languages"
llm -m openai/o3-2025-04-16 "say hi in five languages"
llm -m openai/o4-mini "say hi in five languages"
llm -m openai/o4-mini-2025-04-16 "say hi in five languages"

llm -m openai/o3 'describe image' -a https://static.simonwillison.net/static/2025/two-pelicans.jpg
llm -m openai/o3-2025-04-16 'describe image' -a https://static.simonwillison.net/static/2025/two-pelicans.jpg
llm -m openai/o4-mini 'describe image' -a https://static.simonwillison.net/static/2025/two-pelicans.jpg
llm -m openai/o4-mini-2025-04-16 'describe image' -a https://static.simonwillison.net/static/2025/two-pelicans.jpg

llm -m openai/o3 'invent a cool dog' --schema 'name,age int,bio'
llm -m openai/o3-2025-04-16 'invent a cool dog' --schema 'name,age int,bio'
llm -m openai/o4-mini 'invent a cool dog' --schema 'name,age int,bio'
llm -m openai/o4-mini-2025-04-16 'invent a cool dog' --schema 'name,age int,bio'