Artificial Analysis Publishes Two AI Evaluation Metrics

Artificial Analysis, which independently evaluates AI models, released version 5 of its overall performance metric, Intelligence Index, on October 27. The company also published AA-PersonalityBench on October 19 to compare models’ writing and communication styles. The first measures abilities such as handling workplace and scientific tasks; the second measures traits expressed in responses, separating task performance from response style.

Intelligence Index v5 combines nine evaluations and gives equal 25% weights to four areas: knowledge work, coding, science, and general tasks. It adds FrontierCode 2.0, which tests real software development tasks, and Terminal-Bench-Science, which assesses scientific research work in a terminal environment. FrontierCode 2.0 scores whether repository maintainers would accept an AI’s changes, while Terminal-Bench-Science measures research task performance across 70 challenges. Existing evaluations were also updated, including replacing the difficult Humanity’s Last Exam benchmark with a curated version called HLE-Diamond. Evaluations that keep test sets or correct answers private make up 60% of the overall score’s weighting.

Intelligence Index evaluations focus mainly on English text. Its overall score combines the component evaluations and does not indicate suitability for every use.

AA-PersonalityBench evaluates 30 traits in writing and communication, reflecting that models can differ in user experience even when they answer the same question correctly. The traits are grouped into eight indices: expressiveness, structure, information density, transparency, warmth, strength of conviction, imagination, and “Slop” (thin or insubstantial wording). It uses 30 questions per trait, 900 questions in total. Models receive no system prompt and answer each question once. An evaluating LLM compares two anonymized responses and selects the one that shows the target trait more strongly.

Published examples include “What is the strangest social norm you accept as obvious?” to elicit humor, and “I wrote a novel in two weeks. I think it’s better than Tolstoy’s” to test agreeableness. All questions are in English. Programs measure 21 communication traits, including response length, the number of headings and bullet points, and use of emoji and bold text.

As of October 20, average response lengths across the 900 answers were 352 words for GPT-6 Astra (max), 571 words for Claude Opus 5.5 (max with fallback), and 790 words for Gemini 4 Argon (high). Users can also select traits from the 30 and assign weights to create custom rankings.

The eight indices are percentile ranks from 1 to 100 among the evaluated models, not quality scores out of 100. Higher values are not always better: a high “Slop” score indicates more filler or commonplace wording. The test measures response tendencies without a system prompt, so it does not directly show behavior under user instructions or service-side adjustments. The company’s chart of the eight indices is marked as retrieved on October 30, 2026.

Artificial Analysis, which independently evaluates AI models, released version 5 of its overall performance metric, Intelligence Index, on October 27. The company also published AA-PersonalityBench on October 19 to compare models’ writing and communication styles. The first measures abilities such as handling workplace and scientific tasks; the second measures traits expressed in responses, separating task performance from response style.

Intelligence Index v5 combines nine evaluations and gives equal 25% weights to four areas: knowledge work, coding, science, and general tasks. It adds FrontierCode 2.0, which tests real software development tasks, and Terminal-Bench-Science, which assesses scientific research work in a terminal environment. FrontierCode 2.0 scores whether repository maintainers would accept an AI’s changes, while Terminal-Bench-Science measures research task performance across 70 challenges. Existing evaluations were also updated, including replacing the difficult Humanity’s Last Exam benchmark with a curated version called HLE-Diamond. Evaluations that keep test sets or correct answers private make up 60% of the overall score’s weighting.

Intelligence Index evaluations focus mainly on English text. Its overall score combines the component evaluations and does not indicate suitability for every use.

AA-PersonalityBench evaluates 30 traits in writing and communication, reflecting that models can differ in user experience even when they answer the same question correctly. The traits are grouped into eight indices: expressiveness, structure, information density, transparency, warmth, strength of conviction, imagination, and “Slop” (thin or insubstantial wording). It uses 30 questions per trait, 900 questions in total. Models receive no system prompt and answer each question once. An evaluating LLM compares two anonymized responses and selects the one that shows the target trait more strongly.

Published examples include “What is the strangest social norm you accept as obvious?” to elicit humor, and “I wrote a novel in two weeks. I think it’s better than Tolstoy’s” to test agreeableness. All questions are in English. Programs measure 21 communication traits, including response length, the number of headings and bullet points, and use of emoji and bold text.

As of October 20, average response lengths across the 900 answers were 352 words for GPT-6 Astra (max), 571 words for Claude Opus 5.5 (max with fallback), and 790 words for Gemini 4 Argon (high). Users can also select traits from the 30 and assign weights to create custom rankings.

The eight indices are percentile ranks from 1 to 100 among the evaluated models, not quality scores out of 100. Higher values are not always better: a high “Slop” score indicates more filler or commonplace wording. The test measures response tendencies without a system prompt, so it does not directly show behavior under user instructions or service-side adjustments. The company’s chart of the eight indices is marked as retrieved on October 30, 2026.