The way AI models are judged is changing. For years, academic exams, competition math problems and narrow coding tests decided which model counted as the most capable. Those measures are running out of headroom, and in January 2026 the benchmarking firm Artificial Analysis released version 4.0 of its Intelligence Index, a redesign that shifts the emphasis from test scores to how much useful work a model can actually do.
The stated aim of the new index is to assess models on real-world performance rather than exam results. Scoring was deliberately made harder: where the leading models previously scored above 70 on the index, the same models land around 50 under the new methodology, leaving room to measure future improvement. According to the published rankings at launch, OpenAI’s GPT-5.2 sat at the top, followed by Anthropic’s Claude Opus 4.5 and Google’s Gemini 3 Pro. The more consequential change, though, is in what the models are being tested on.
Out with academic tests, in with occupational tasks
The biggest shift is the removal of several well-known academic benchmarks, including school-style math and coding tests, in favor of evaluations tied to economic value. The clearest example is GDPval-AA, an adaptation of OpenAI’s GDPval, which assigns models tasks drawn from 44 occupations across nine major sectors of the US economy. Instead of short answers, models must produce complete deliverables such as reports, spreadsheets and professional documents, on tasks of the kind usually done by workers with many years of experience.
Research-level science and a hallucination test
For scientific capability, the index adds CritPt, a benchmark developed by more than 50 working physics researchers. It consists of unpublished, research-level problems spanning eleven subfields of physics. Even the most advanced models solved only around 11 percent of them, according to the results reported with the index, a sign of how far AI remains from independent scientific reasoning.
Another notable addition is AA-Omniscience, which targets hallucination, the tendency of models to state falsehoods confidently. It poses 6,000 questions across 42 topics and penalizes wrong answers while rewarding a model that says it does not know. Artificial Analysis reported that only three models achieved a positive score, with most systems more likely to guess wrongly than to admit uncertainty.
Why it matters
The redesign arrives as competition among the major AI labs has intensified, with each new release putting pressure on rivals. Companies increasingly need to demonstrate reliability and practical value, not just raw intelligence, and a benchmark built around work tasks pushes them to optimize models for real problems rather than for leaderboard scores.
The new index is composed of four equally weighted areas: agents, coding, scientific reasoning and general knowledge. Artificial Analysis says that after repeated runs the margin of error in a model’s score is below one percentage point, which it presents as evidence that the measure is stable. Artificial Analysis has continued to revise the index in the months since, adding further agentic evaluations.
Taken together, the changes mark a new phase for the industry. The question is no longer which model scores best on an exam, but which one proves dependable, useful and economically valuable in practice, and that is the measure likely to shape AI development priorities in the years ahead.
