@OpenAI Decisions API is multimodal. So you can stuff text into images and save about 20% tokens.

But unfortunately this results in about 13% lower accuracy on tasks on the tests I ran (SQL query classification)