OpenAI vision bug fix transforms GPT-6 Sol and Luna image results
On September 25, OpenAI slipped a vision bug fix into GPT-6 Sol and Luna. The change was buried in an API changelog. But the effect was huge. Teams that wrote off these models for failing image tasks now face a new reality. The fix changes how Sol and Luna read images. Any workflow that relies on screenshots, scanned docs, charts, or browser UIs is affected. Old conclusions may not hold.
Prior to the patch, OpenAI recommended that users double-check any workflow where Sol or Luna analyzed screenshots, scans, charts, or interfaces, as these scenarios were most sensitive to the bug.
What now for teams who ran early tests? If an image task failed after Sol and Luna launched on September 22, the bug may be to blame-not the model's real limits. Rerunning every test is wasteful. But skipping the fix means you might toss out a tool that now works. The smart move is to replay only image-dependent cases. Use the same image files, request setup, and grading rules. Compare before and after. That's the only way to see if the fix, not a cleaner prompt, made the difference.
Text-only tasks are safe. This bug only broke input understanding. It did not affect image generation. If your workflow just creates images with another model, you're fine. But if you send those images back into Sol or Luna for review, editing, or computer control, the fix matters. The most urgent retests are high-stakes ones: document extraction that moves data, UI automation that can submit or delete records, visual checks that approve assets, or any test that led you to reject Sol or Luna for a job. Start with a small set. Include both past failures and wins. Otherwise, you won't know if the model's accuracy changed or if only a few edge cases got better.
OpenAI recommends rerunning evaluations and workflows that use image inputs, as the bug affected image understanding but not image generation. The company has emphasized that new requests to the API and Codex now use the corrected image processing.
OpenAI's silence on the details leaves teams with one clear rule. Treat September 25 as a hard line in your reports. Keep pre-fix results for the record. Run new tests on the same inputs. Note any other changes. If results get better, update your baseline only after full checks pass. If not, at least you know what to fix next-image quality, prompt, tool setup, or a model limit that's still there.
The patch is not a magic cure. One vendor chart does not prove the problem is gone. But early image failures now need a second look. In AI testing, trust comes from replaying real work. Not from a patch note. Not from a marketing slide. Teams that retest their key visual workflows now will know first if Sol and Luna finally deliver-or if the hunt for reliable vision AI goes on.
