An image-encoding bug was degrading visual understanding for GPT-6 Sol and Luna in the API and Codex, including computer-use workflows. OpenAI’s September 25 changelog describes a fix to that processing path, not the release of a new model.
The provider recommends rerunning affected evaluations and workflows that use images. A before-and-after improvement may therefore reflect the encoding correction rather than only a change in model generation. That matters when interpreting visual comparisons: results obtained on the faulty path and after the fix do not describe identical conditions.
