Skip to content
Brief The AI-First Web ·

GPT-6.2 models get to the point faster but rank lower on admitting doubt, test finds

An Artificial Analysis benchmark ranks OpenAI's Astra and Sol below two Claude models on transparency.

In brief
  • Astra and Sol write answers 34 to 40 percent shorter than their predecessors. Their Transparency Index ranks are 41 and 57.
  • The scores rank models against each other and do not measure accuracy. The Decoder says a lower score does not necessarily mean more wrong answers.

OpenAI launched GPT-6.2 Astra and Sol on 17 November. The Decoder reported on 22 November that Artificial Analysis, which tests AI models, found their answers much shorter. Astra's average dropped from 352 to 210 words, and Sol's from 337 to 223. Both also lead with the point more often. Astra's rank for answering first rose from 49 to 89, and Sol's from 64 to 87.

AI judges compare two anonymous answers and pick which shows a trait more, or call a tie. Results become ranks from 1 to 100 among tested models, not shares of answers. Per The Decoder, Astra's rank for admitting uncertainty fell from 96 to 24, and Sol's from 98 to 51. Astra scores 41 and Sol 57 on the Transparency Index, against 81 and 90 for Anthropic's Claude Fable 5.2 and Opus 5.5. That index blends five traits, so doubt is only one part. The comparison uses four "max" variants and 900 English questions asked without a system prompt. A lower score does not necessarily mean more wrong answers.

OpenAI's launch post stresses fewer output tokens, and the test ranks these models lower on admitting doubt. The sources do not say one causes the other. Buyers can run their own prompts, include some with no clear answer, and see whether the model says so. They can also test with their own system prompt, and ask vendors how they handle doubt.

A WebPulse Brief: a short report of an important event, written by the WebPulse Newsroom with AI assistance and checked against the reporting below. When there is more to explain, we follow up with a full story. How we use AI.

Reporting: Artificial Analysis, OpenAI, Simon Willison.