GPT-6.2 Astra flubs "latest Claude model" question, only 45% correct

bdsqlsz · x · 2026-11-18

Developer bdsqlsz reports serious hallucination issues in GPT-6.2 Astra: asking the same question repeatedly under the same IP and environment yielded only 45% correct answers. Screenshots show the model naming "Claude Opus 4.7 (April 2026)" in one run and a seemingly fabricated "Claude Fable 5, announced alongside Claude Mythos 5 in June 2026" in another, with response times ranging from 36 seconds to nearly 12 minutes. The author shared the test prompt — asking the model to name the latest Claude model using offline knowledge only — as a quick probe of knowledge cutoff and hallucination.

Related event: GPT-6.2 Astra Names the Latest Claude Model Correctly Only 45% of the Time(2 posts)→

Original post →

More from Fun

Fun channel →