Artificial Analysis launches VIALS-AA, a life-sciences visual-reasoning leaderboard where Gemini 4 Carbon leads at 64%

Artificial Analysis has launched VIALS-AA, a leaderboard for Handshake AI's VIALS benchmark. VIALS has 161 tasks in eight domains, in which a model must read a scientific artifact (gel blots, flow cytometry plots, plasmid maps, phylogenies) and draw a specific conclusion.

Artificial Analysis runs each task three times in its open-source Stirrup agent harness, with code execution and image inspection tools. GPT-6 Luna (medium) grades each answer against the reference, and the score is the share of tasks judged correct in all three runs.

According to Artificial Analysis:

• Top scores: Gemini 4 Carbon (high) leads with 64%, ahead of GPT-6 Astra (max) at 63% and Claude Opus 5.5 (max with fallback) at 61%.
• Cost: GPT-6 Astra and GPT-6.1 Sol (both max) use about 8 turns per task, against about 27 for Claude Opus 5.5 and about 25 for Claude Sonnet 5.5. Opus 5.5 averages $1.64 per task, about 29 times Sol's roughly $0.056.
• Open weights: Kimi K3.1 Pro leads at 45%, narrowly ahead of MiMo-V2.6-Pro at 43%, about 19 points behind the top score.
• Refusals: 14 of the 161 tasks drew a safety refusal from at least one model; Claude Sonnet 5.5 refused on 6, the most of any model.

These are Artificial Analysis's own measurements.