OpenAI Fixes GPT-6 Image Bug, Luna Visual Localization Doubles

Beating AI News Flash: OpenAI has fixed an image encoding bug in GPT-6 Sol and GPT-6 Luna. The issue reduced the models' ability to understand images. After the fix, vision tasks in both the API and Codex will improve, including Computer Use, which allows Agents to look at the screen and operate a computer.

In tests published by OpenAI, GPT-6 Luna (Max)'s average IoU on the RefCOCOg visual grounding benchmark rose from 28.8% to 60.0%, an increase of 31.2 percentage points. This test mainly measures whether a model can accurately locate the corresponding object or region in an image based on a single description.

The two models only officially launched their APIs on September 22, and OpenAI's changelog recorded this fix on September 25. The company did not disclose the specific improvement for Sol, nor did it explain when the bug began or how many image requests were affected.

OpenAI also rarely advised developers using image inputs to rerun their evaluations and retry previously affected workflows.