OpenAI retunes GPT-5.6 Sol and gives free users unlimited Luna chats
The Sol retune cuts factual errors 68%; Plus and Pro get a slider to dial reasoning effort, and the free tier flips to Luna as default with a new Think button.
Copy markdown
68% fewer wrong answers
OpenAI says GPT-5.6 Sol makes about 68% fewer factual errors than 5.5 Instant on financial, medical, and legal prompts (Luna is ~62% fewer), with tighter formatting and better use of sources. If you'd been dropping to Instant for speed, that quality gap just narrowed.
A slider for how hard it thinks
Plus and Pro now get a slider on web, mobile, and desktop to set how much thought ChatGPT spends per answer -- a manual latency-vs-quality dial instead of picking a model by hand. Sol also folds the old Instant/Thinking split into one model.
Free tier flips to Luna, unlimited
Free and Go users get GPT-5.6 Luna as the default this week, with unlimited text chats and a new Think button for harder questions. File uploads, images, and other tools stay capped, but the daily text-message wall is gone.
Trained to back off with teens
The models were tuned to avoid romantic roleplay, age-restricted challenges, and posing as a substitute for real relationships, plus firmer limits on self-harm, eating-disorder, and graphic-violence content. Expect more refusals on those edges regardless of tier.
Elsewhere: a cheap open reasoning model to try
inclusionAI's Ling 3.0 Flash -- a 124B MoE with only ~5.1B active params, 256K context, MIT-licensed -- is now broadly available on OpenRouter and DeepInfra after its free promo ended. It posts AIME 2026 93.2% and SWE-Bench Pro 56.6%, and runs locally in Ollama or LM Studio.