OpenAI updates GPT-5.6 Sol, cuts factual errors, frees Luna

What happened: OpenAI released an updated GPT-5.6 Sol in ChatGPT on 6 August 2026, tuned for shorter, more direct answers that correct users rather than simply agreeing with them.

The numbers: In OpenAI's own tests on financial, medical and legal prompts, responses containing at least one factual error were 68% less common with GPT-5.6 Sol and about 62% less common with GPT-5.6 Luna, compared with GPT-5.5 Instant. OpenAI credits better use of the sources it looks up.

The catch: This is OpenAI's own testing, not an independent benchmark. The update is chat-only: Sol in ChatGPT Work and Codex is unchanged, and this is still GPT-5.6, not a new model generation.

Who's affected: Plus and Pro users get a single Sol model that now covers both quick answers and deep reasoning, controlled by a slider. Free and Go users get a new "Think" button instead.

The details: Free and Go users also now get GPT-5.6 Luna as their default model, with unlimited text chats rolling out from 7 August.

Why it matters: Anthropic currently leads business spending on AI. OpenAI is betting on better answer quality and wider free access to hold onto its consumer lead.

The context: Protections for under-18s were also strengthened, including new limits on romantic role-play.

What's next: OpenAI moved the model picker into settings, a change some users have already criticised on X. More detail from DataNorth.