OpenAI has changed how ChatGPT talks to most of the people who use it. From 29 October, the free version gives shorter answers with less flattery. OpenAI’s own tests show a cost: the updated model made more factual errors than the version it replaced, on all six accuracy measures the company reported.
Who gets the change
According to OpenAI’s model release notes, the update applies to Free and Go users and to anyone using ChatGPT without signing in. On the Plus and Pro plans it only affects the backup model ChatGPT switches to once you hit your GPT-6 message limit. The main GPT-6 model on those plans is unchanged.
OpenAI says the aim is “more concise responses” and less “overly flattering language”. Flattery here means a chatbot that praises your ideas or agrees with you too readily, even when you are wrong.
The numbers
The detail is in a system card addendum, the safety report OpenAI publishes with each model update. On HealthBench, OpenAI’s test set of health questions, the average answer fell from 3,361 to 2,970 characters. That is about 12% shorter. The health score, adjusted for length, slipped from 50.4 to 48.2 out of 100.
On accuracy, OpenAI says the new version “shows higher factual-error rates” than the 7 October version across all six metrics it reported, though it still makes fewer errors than the older GPT-5.6 Luna. The addendum shows those error rates only as a chart, not as figures in the text.
There were gains too. On a test of how the model responds when users under 18 lean on it for emotional support, the score rose from 0.734 to 0.870, where 1 is best. OpenAI also reports better scores for teenagers on sexual content, eating disorders and gore.
Why it matters
Flattery has been a long-running problem for ChatGPT. In April 2025 OpenAI’s chief executive Sam Altman admitted its personality had become annoying and promised a fix. A chatbot that tells people what they want to hear is pleasant, but it can be harmful when someone asks about their health, money or a big decision.
On the same day OpenAI also rewrote its Model Spec, the public rulebook for how its models should behave. It now rests on five values, including honesty and “human agency”, meaning people should be helped to make their own choices rather than steered.
For people on the free version, the practical upshot is simple. Expect shorter, blunter answers. By OpenAI’s own measure they are also a little less reliable, so check anything that matters against a second source.
Get AI to do the boring stuff
Our free AI agent playbook: simple, step-by-step recipes anyone can follow. Plus one short email a day with the AI story that matters.
Source link