OpenAI fixes a bug that degraded image understanding in GPT-6 Sol and GPT-6 Luna

OpenAI has fixed a bug in image encoding that degraded image understanding in GPT-6 Sol and GPT-6 Luna, and its changelog dates the fix 25 September. If you send images to either model, what you measured before that day may not be what you get now — and OpenAI itself asks you to measure again.

What changed

The fix improves results on visual tasks in the API and in Codex, including computer use. The entry is marked as a fix and names the two model IDs it touches,gpt-6-sol and gpt-6-luna. It names no new model, no new parameter and no change to how you call either one: the requests you send today are the ones you sent before.

What it was before

Both models were released on 22 September, three days before the fix. They are reasoning models that accept text and image inputs and generate text through the Responses and Chat Completions APIs , and each model page lists text and image as input and text as output.

The changelog does not say when the bug began. What the two dates do set is the span in which it could matter: from the release on 22 September to the fix on 25 September, any image either model read may have been read worse than it is now.

What it means for you

OpenAI's advice is direct: if your use cases involve image inputs, rerun your evaluations and retry the workflows the issue affected. In practice that means three places to look.

Anything that reads images.OpenAI's vision guide lists what image understanding covers: describing images, reading visible text, and answering questions about objects, shapes, colors or textures. If you tested any of those on Sol or Luna before 25 September — reading receipts, checking screenshots, describing product photos — test again before you trust the old numbers.

Computer use.In computer use the model decides what to do next from screenshots and other tool results , which is why the fix names it. The computer use guide's own code example defaults togpt-6-astra, and the fix does not name Astra. If you ran computer use on Sol or Luna instead, those runs are the ones to repeat.

A choice of model you already made. If you tried Sol or Luna on an image task that week, saw it do poorly and moved to another model, that comparison was run against a model with the bug. Run it again; the answer may change.

Text-only work is not part of this. The entry names image understanding and nothing else.

If you are new to these two models, our news item on their release sets out what each is for.

The source

OpenAI's changelog, the entry dated 25 September 2026, marked Fix.