Understand this story
On November 6, 2026, Artificial Analysis launched the VIALS-AA scientific image reasoning evaluation. The evaluation uses 161 tasks, running each task 3 times, to test models' ability to interpret scientific images. Results show that Gemini 4 Carbon (high) ranks first with a pass^3 score of 64%, answering 2 more tasks correctly than GPT-6 Astra (max). Claude Sonnet 5.5 (max, with fallback) and GPT-6.1 Sol (max) scored closely, at 60% and 59% respectively, but the former made 3.6 times as many tool calls per task as the latter. Per-task cost ranged from $0.02 for GPT-6 Luna (max) to $1.77 for Grok 4.7 (xhigh), a difference of nearly 99 times.
AI-generated from public reports
Latest development · Report
VIALS-AA: Different Paths to Scientific Image ReasoningDevelopments
1 confirmed developments
Artificial Analysis launches the VIALS-AA scientific image reasoning benchmark
Source:Artificial Analysis (full article)
Reports and discussion timeline
Browse reports and discussion by time. Their evidence levels remain distinct.
Nov 6
Artificial Analysis full articleFirst-party
VIALS-AA: Different Paths to Scientific Image Reasoning
Artificial Analysis has released the VIALS-AA benchmark, running each of 161 tasks 3 times to test models' ability to interpret scientific images. Gemini 4 Carbon (high) tops the list with a pass^3 score of 64%, answering 2 more tasks correctly than GPT-6 Astra (max). Claude Sonnet 5.5 (max, with fallback) and GPT-6.1 Sol (max) score closely (60% and 59%), but the former makes 3.6 times as many tool calls per task as the latter. Per-task cost ranges from $0.02 for GPT-6 Luna (max) to $1.77 for Grok 4.7 (xhigh), a difference of nearly 99 times.