Initial plugin design

[simonw]

LLM's OpenAI functionality is currently bundled with LLM itself in https://github.com/simonw/llm/blob/0.23/llm/default_plugins/openai_models.py

This means that any time OpenAI release a new model or feature I have to push a new release of LLM itself!

I also need to rework the OpenAI support for the new Responses API. I'm using that as an excuse to split out the code into a fresh plugin.

[simonw]

For the first release this will register models as openai/gpt-4o-mini etc - with that openai/ prefix. That will allow the plugin to be installed at the same time as the current default plugin.

Later I'll remove the default plugin from LLM core and use this one instead. At that point I'll add aliases for the old gpt-4o-mini model names.

[simonw]

Interesting - I tried this:

llm -m openai/gpt-4o-mini describe \
  -a https://static.simonwillison.net/static/2025/two-pelicans.jpg

And then llm logs --json -c and saw this:

{
  "usage": {
    "input_tokens": 37073,
    "input_tokens_details": {
      "cached_tokens": 0
    },
    "output_tokens": 83,
    "output_tokens_details": {
      "reasoning_tokens": 0
    },
    "total_tokens": 37156
  },
  "user": null,
  "store": true
}

Why so many tokens?

https://platform.openai.com/docs/guides/images?api-mode=responses&format=url#specify-image-input-detail-level says

The detail parameter tells the model what level of detail to use when processing and understanding the image (low, high, or auto to let the model decide). If you skip the parameter, the model will use auto. [...]

You can save tokens and speed up responses by using "detail": "low". This lets the model process the image with a budget of 85 tokens. The model receives a low-resolution 512px x 512px version of the image. This is fine if your use case doesn't require the model to see with high-resolution detail (for example, if you're asking about the dominant shape or color in the image).

I tried adding "detail": "low" and ran it again and got:

{
  "usage": {
      "input_tokens": 3071,
      "input_tokens_details": {
        "cached_tokens": 0
      },
      "output_tokens": 76,
      "output_tokens_details": {
        "reasoning_tokens": 0
      },
      "total_tokens": 3147
   }
}

WAY cheaper.

[simonw]

So I guess I need a new option something like -o image_detail high - I think I'll default it to "low" to avoid expensive surprises.

[simonw]

Rats, need to find a different name. Probably caused by https://pypi.org/project/llmopenai/

[simonw]

I'm going with llm-openai-plugin for the moment, but I'll keep the module itself as llm_openai.

[simonw]

First alpha is live: https://pypi.org/project/llm-openai-plugin/