Free ChatGPT Version Now Has More Errors, Less Flattery
OpenAI's free ChatGPT update cuts flattery but raises factual error rates; verify important answers.
What the Numbers Show
Who Should Care
The Takeaway
OpenAI changed the default behavior of free ChatGPT on 29 October, swapping in a model that produces shorter, less flattering replies. The catch, according to OpenAI's own system card addendum, is that the new version also made more factual errors than the 7 October model it replaced across all six accuracy metrics the company reported. The update hits Free and Go users, people who use ChatGPT without signing in, and Plus/Pro subscribers only after they exhaust their GPT-6 message allowance; the main GPT-6 model on paid plans is unchanged.
## What the Numbers Show
On HealthBench, the average health answer dropped from 3,361 to 2,970 characters, about 12% shorter, while the length-adjusted health score slipped from 50.4 to 48.2 out of 100. OpenAI released the error-rate increases as a chart rather than exact figures, but it did note the model still makes fewer errors than older GPT-5.6 Luna. Some safety measures improved: emotional-support scores for users under 18 rose from 0.734 to 0.870, and teen safety scores improved on sexual content, eating disorders, and gore. So this is not a simple downgrade—it is a deliberate trade of factual precision for brevity and better safety behavior.
## Who Should Care
Anyone using free ChatGPT, the ChatGPT app without an account, or the backup model on paid plans is now absorbing that trade-off. For casual questions, shorter answers are likely fine. For medical, financial, legal, or other high-stakes decisions, the higher error rate means you should verify anything that matters against a second source. OpenAI's rewritten Model Spec also emphasizes honesty and human agency, which explains the reduced flattery, but less praise does not automatically mean more accuracy.
## The Takeaway
A chatbot that stops telling you what you want to hear is progress, particularly for vulnerable users. But because OpenAI itself flags rising factual errors, treat free ChatGPT as a useful draft generator—not a reliable authority—and double-check critical information.
Use free ChatGPT as a quick draft tool, not a final source, until accuracy improves.