The new GPT-5.6 Sol powers all chats for paid users, including Instant, creating

OpenAI's 68% figure is a relative comparison from its own test, not an accuracy rate. In a two-post thread on August 6, 2026, OpenAI said that on its high-stakes factuality evaluation covering finance, medicine and law, the updated GPT-5.6 Sol produced 68% fewer responses with factual errors than GPT-5.5 Instant. The thread's text does not give either model's actual error rate, the number of prompts or how errors were judged. A reader can conclude that OpenAI measured Sol as making errors in far fewer answers on that test, but not how often either model is wrong.

The figure comes with a change to which model answers. For Plus and Pro subscribers, GPT-5.6 Sol now powers both the Instant setting and deep reasoning, which OpenAI describes as one consistent experience. Free and Go users instead get unlimited text chats with GPT-5.6 Luna starting August 7, the day after the announcement, and the thread gives no factuality figure for Luna. Because the baseline is GPT-5.5 Instant, the number also does not show how the updated Sol compares with the earlier version of Sol. It is OpenAI's measurement on its own evaluation, not independent evidence of an error rate for everyday questions.