I ended up using the HTTP Request node. So far, so good.
FYI – gpt-4-vision-preview sometimes answers “Sorry, I cannot help with you that” (or similar) on the first pass on a small % of images for which I prompted it for a description (maybe 3%). It’s not related to TPD, RPD, TPM API limits because I paced it to run very slowly. Looks like a second pass might work just fine.