TL;DR

OpenAI's September 25 changelog reports an image-encoding fix affecting GPT-6 Sol and Luna in visual workflows. Earlier image-input failures may therefore deserve a fresh evaluation even if the model name is unchanged. Keep the run date and conditions with the result, and rerun representative cases. The notice does not quantify a universal accuracy improvement.

What changed

OpenAI reports fixing an image-encoding bug that impaired image understanding in GPT-6 Sol and GPT-6 Luna. Its changelog recommends rerunning image-input evaluations and retrying affected workflows. It does not quantify a universal improvement.

What it means for your workflow

SaveMyToken editorial take

Keep the run date with a visual-task evaluation. An earlier failure may need a fresh check after a serving fix, even when the model name stays the same.