OpenAI says it fixed, on Sept. 25, a bug in the step that prepares submitted images for GPT-6 Sol and GPT-6 Luna to interpret. The bug had degraded the models’ image understanding. The company recommends that developers using image inputs rerun their evaluations and retry workflows affected by the issue.
OpenAI’s API changelog says the fix improves visual tasks in the API and Codex, including computer use. The company has not identified the faulty operation, the affected time window or which requests were affected, limiting what developers can infer without testing their own applications.
Andon Labs said on Sept. 27 that it reran its Blueprint-Bench 2 evaluation after the fix. The benchmark asks an agent to turn about 20 interior photos from each of 50 apartments into floor plans showing room layout, connections and relative sizes. Andon reported a small gain for Sol and a much larger one for Luna, but did not publish before-and-after scores in its post.
The benchmark’s current table lists GPT-6 Sol at 0.369 and Luna at 0.312 from runs that each process 50 apartments. Its connectivity score combines several measures, with room-to-room connection overlap carrying half the weight; scores are normalized so that a random baseline is 0 and a perfect result is 1. The human comparison of 0.586 comes from a separate 12-apartment subset. The page does not provide a versioned before-and-after table, so these numbers describe the listed runs, not the improvement caused by the fix.
This is evidence from one image-heavy spatial task, not a measure of all visual recognition or computer-use workflows. Teams using image inputs should follow OpenAI’s advice and rerun evaluations built around their own tasks, particularly when checking a baseline that may have included the bug.
Photo: Javan Rhino, Homo sapien using laptop at night, CC BY 4.0. No changes made.
