GPT-6.2 Answers Are 40% Shorter, and It Almost Never Admits It Doesn't Know Anymore
Artificial Analysis's PersonalityBench shows that GPT-6.2's average response length fell from 352 words to 210 words, and its score for admitting uncertainty dropped from 96 to 24. Shorter answers look more confident, but confidence does not equal correctness.
Artificial Analysis released the PersonalityBench scores on November 22. GPT-6.2 (test codename Astra) has become shorter: its average response dropped from 352 words to 210 words, a 40% decrease. At the same time, it also rarely says “I don't know” anymore.
PersonalityBench measures the style of model responses and does not directly represent capability. In the Admits Uncertainty item, GPT-6.2's score fell from 96 to 24. On another metric, the Transparency Index, Astra scored 41; Claude Fable 5.2 scored 81; Opus 5.5 scored 90.
Short answers look decisive. But decisiveness does not equal correctness. After seeing this set of data, some online commenters said: the drop from 96 to 24, paired with the comparison of transparency index scores, is indeed striking.
This round of evaluation did not test capability; it tested style. But style is also part of the product. When models learn to express confidence in shorter sentences, users need to be more careful: it can answer more elegantly, but it can also be wrong in more hidden ways.
One clear signal is that the model's expression of uncertainty is being compressed. A score drop from 96 to 24 means that when it encounters questions it is unsure about, it is less likely to admit it doesn't know. When developers evaluate models, in addition to whether responses are fluent, they also need to check whether the model honestly expresses confidence.
发布时间: 2026-11-22 23:32