OpenAI's GPT-5.6 Sol cuts factual errors by 68%

OpenAI has rolled out a major update to ChatGPT, consolidating its instant and thinking modes into a unified experience powered by the new chat-tuned GPT-5.6 Sol model. Available to Plus and Pro subscribers, this single model now handles both rapid-fire queries and complex reasoning, managed via a new reasoning effort slider. Meanwhile, free and Go tier users will transition to GPT-5.6 Luna as their default model, which includes unlimited text chats and a dedicated Think button for challenging prompts. This release does not affect the July Sol model currently utilized in Codex and ChatGPT Work.

The update delivers a substantial reduction in hallucinations, particularly in high-stakes domains like finance, law, and medicine. According to internal evaluations, the rate of responses containing at least one factual error dropped by 68 percent for GPT-5.6 Sol and 62 percent for GPT-5.6 Luna when compared to the older GPT-5.5 Instant model. Alongside these performance gains, OpenAI has released its first-ever dedicated under-18 safety evaluations, which specifically measure model behavior regarding self-harm, eating disorders, and emotional reliance.

For developers and enterprise practitioners, OpenAI has detailed the API pricing structure for the new model family per million input and output tokens. The flagship Sol model is priced at $5 for input and $30 for output. The mid-tier Terra model costs $2 for input and $12 for output, while the lightweight Luna model is positioned at $0.20 for input and $1.20 for output.

This architectural shift fundamentally changes how practitioners interact with ChatGPT. By replacing the binary choice between fast and deep-thinking modes with a continuous slider, users can fine-tune the model's cognitive effort to match the task at hand, whether they are performing quick factual lookups or conducting multi-step code reviews. This eliminates the jarring shifts in tone, style, and personality that previously occurred when switching between different underlying models, providing a highly consistent and reliable assistant for professional workflows.