OpenAI fixes GPT-6 image bug, Luna's visual positioning accuracy doubles

Insight: Beating AI Flash News – OpenAI has fixed an image encoding bug in GPT-6 Sol and GPT-6 Luna. The bug was impairing the models’ image comprehension capabilities; following the fix, visual tasks on both the API and Codex will be enhanced, including the Computer Use feature that enables agents to view screens and operate computers. In tests released by OpenAI, GPT-6 Luna (Max) saw its average Intersection over Union (IoU) on the RefCOCOg visual grounding benchmark rise from 28.8% to 60.0%, a 31.2 percentage point improvement. This test measures a model’s ability to accurately locate the corresponding object or region in an image based on a textual description. The two models only officially launched their APIs on September 22, and OpenAI documented the fix in its September 25 update log. The company did not share specific performance gains for Sol, nor did it disclose when the bug originated or how many image requests were affected. OpenAI also unusually advised developers using image inputs to rerun their evaluations and retry previously impacted workflows.