OpenAI cuts factual errors by 68% in GPT-5.6 Sol and removes chat limits for free users

OpenAI is rolling out a pair of model updates that feel less like a flashy launch and more like the company finally sanding down the rough edges that have annoyed power users for months. The flagship upgrade is GPT-5.6 Sol, which is now the standard for Plus and Pro subscribers. The headline number is a 68% reduction in factual errors on financial, medical, and legal prompts compared to the old GPT-5.5 Instant model. That’s not incremental improvement—that’s the kind of reliability leap that makes you reconsider using ChatGPT for work you’d previously have triple-checked manually.

The company is also collapsing the awkward split between its “Instant” and “Thinking” modes. Instead of feeling like you’re switching between two different AI personalities, a new slider lets users dial in how much reasoning time the model gets. Quick question? Keep it short. Multi-step planning or research? Crank it up. The model now adapts its verbosity and formatting to the task, which means fewer walls of bullet points for simple queries. As OpenAI puts it, the model will even offer a “helpful correction when simply agreeing wouldn’t be useful.” That’s a diplomatic way of saying it’s been trained to push back on your bad ideas.

On the free tier, the updates are arguably more significant. GPT-5.6 Luna becomes the default model, and OpenAI is removing rate limits on text chats entirely. Free users also get a “Think” button that gives Luna extra compute for harder questions—a feature previously locked behind a paywall. With 1 billion weekly users, the economics of serving unlimited chats to the free tier must be eye-watering, but it signals OpenAI is prioritizing habit formation and market share over immediate margins. This is the company betting that ubiquity now pays dividends later.

There’s a quiet but meaningful safety section buried in the announcement. For users flagged as under 18, the model has been explicitly trained to avoid romantic roleplay, refuse age-restricted challenges, and redirect teens toward trusted adults when they need support. It’s a recognition that ChatGPT has become a de facto companion for a lot of young people, and a guardrail set that competitors like Character.AI are still wrestling with. The model also refuses to present itself as a substitute for real-world relationships, which is a telling line—it suggests OpenAI sees the emotional attachment problem clearly and is building defenses against it before regulation forces their hand.