Top score on AA-Omniscience, a test that punishes made-up answers, rose from 4.8 to 67 in a year
Artificial Analysis scores facts recalled minus facts invented. Claude Fable 5.2 leads; two Gemini 4 models rarely guess wrong.
AA-Omniscience asks models factual questions from six fields: business, humanities, science, health, law and software engineering. A right answer adds points, a wrong one subtracts points and declining to answer costs nothing. The Net Score runs from -100 to 100, and 0 means as many right answers as wrong ones. Artificial Analysis runs every model itself, so the scores are independent of the labs.
At launch in November 2025, Claude 4.1 Opus led with 4.8, and only three models scored above zero. GPT-5 (high) and Grok 4 had the best accuracy, 39%, but gave wrong answers for 81% and 64% of the questions they did not get right.
As of 10 November 2026 the page shows Claude Fable 5.2 (max) at a Net Score of 67 and an accuracy of 79%. The lowest hallucination rates belong to Gemini 4 Carbon (Medium) and Gemini 4 Argon (High) at 4%, then Gemini 4 Carbon (High) at 11%. The page defines the rate as incorrect answers divided by all non-correct responses (incorrect, partial and declined). A 4% rate therefore means that when the model did not know, it said so or hedged 96% of the time. It does not mean it is wrong 4% of the time.
What it means
Getting more right and inventing less are separate skills, and different models lead on each. Neither of the two Gemini models with the lowest rate is among the top three on Net Score or accuracy, while Claude Fable 5.2 holds all three of those places. A reader who cares about wrong answers more than missing ones should look at the hallucination rate; one who wants the most facts recalled should look at accuracy.
Caveats
- The two ends of the headline come from different question sets. The launch score is on v1 (6,000 questions). Today’s scores are on v1.1 (3,000 questions, with saturated questions removed). The page’s chart of v1 against v1.1 scores puts the two versions 8.9 points apart for Claude Fable 5.2 (max), small next to a climb of about 62 points.
- Hallucination rate is a share of the questions a model did not get right. A model that gets 79% right has few left, so the rate rests on a smaller base than accuracy does.
- Each question is asked once. The questions are written by an LLM agent from published sources and test recall without search tools, so the scores do not describe a model that can look things up.
- Artificial Analysis keeps the v1.1 question set private and publishes a public sample. It reports that public and private scores agree closely across 86 models (r = 0.991).
- The page’s text lists the top three models per measure. Other models’ accuracy and hallucination rates appear only in its charts.