A request with an inline image no longer holds tokens for its base64
Before a request runs, Inference Bureau holds an upper bound of its input tokens against rate limits and budgets, counted from the body's length. An image sent inline as base64 was counted as text, so a 300 KB screenshot held about 200,000 tokens and could be refused by a tokens-per-minute limit or a model's context size, though models bill an image by its pixels: a few thousand tokens at most.
Each inline image is now held for at most 8,192 tokens, in chat completions (image_url with a data URL), Messages (an image block's base64 source) and Responses (input_image) bodies, tool results included. What a request settles at is unchanged: the tokens the provider billed. Images passed by address, documents and everything else are held as before.