An image encoding error that hampered visual tasks in the API and Codex has been resolved for the GPT-6 Sol and Luna models. The patch restores the models' ability to process visual data and execute computer use functions, correcting a bug that caused performance in GPT-6 Luna to fall below terms of the earlier GPT 5.6 version. Benchmarks for high-difficulty text extraction in Luna rose from 66.67% to 84.88%, while visual reasoning scores climbed from 60.71% to 71.08%.

The technical update also improved efficiency for GPT-6 Luna, reducing average latency by 25% and estimated costs by 17% while cutting output tokens by 20%. OpenAI released the fix over a weekend following public reports and benchmarking that identified the deficiency. The company has instructed developers to test the restored capabilities in the API and Codex immediately.