OpenAI's GPT-5.6 Sol Cuts Hallucinations 68% and Ditches Thinking Modes
OpenAI upgrades GPT-5.6 Sol for paid ChatGPT users with 68% fewer factual errors and gives free users unlimited chats with GPT-5.6 Luna plus a new Think button
- GPT-5.6 Sol (chat-tuned) is now the single model powering both Instant and reasoning for Plus/Pro ChatGPT users, with a new reasoning effort slider.
- Factual error rate drops 68% vs GPT-5.5 Instant on high-stakes finance, medicine, and legal prompts; 62% improvement for Luna.
- Free and Go users get GPT-5.6 Luna as default with unlimited text chats and a new Think button for harder questions.
- This chat-tuned Sol is separate from the July Sol used in Codex and ChatGPT Work — those are unchanged.
- OpenAI published first-ever dedicated U18 safety evaluations alongside this release, covering eating disorders, self-harm, and emotional reliance.
- API pricing: Sol at $5/$30, Terra at $2/$12, Luna at $0.20/$1.20 per million input/output tokens — full announcement here.
OpenAI has updated ChatGPT in a way that retires the old split between "Instant" and "Thinking" modes. For Plus and Pro users, a freshly tuned version of GPT-5.6 Sol now handles everything from quick replies to deep reasoning in a single model. For free users, GPT-5.6 Luna becomes the new default, and text chats are now unlimited. The goal is a frontier model that behaves like one coherent assistant rather than a collection of modes stitched together.
One model, one experience
Previously, switching from a fast Instant response to a Thinking response could feel like talking to a different model. With this update, the same model powers both fast replies and deeper reasoning for Plus and Pro users, so tone, style, and personality stay consistent whether you ask a quick question or kick off a multi-step research task.
A new slider lets you choose how much reasoning ChatGPT applies to each response. Keep it low for quick factual lookups, nudge it up for planning, writing, or code review. This replaces the binary Instant/Thinking toggle with a continuous dial.
Fewer hallucinations where it matters
The headline improvement is factual accuracy. In an internal evaluation of financial, medical, and legal prompts, responses containing at least one factual error were about 62% less common with GPT-5.6 Luna and 68% less common with GPT-5.6 Sol compared to GPT-5.5 Instant. OpenAI's