Artificial Analysis launches VIALS-AA to test how reliably AI models read scientific images

On November 6, 2026, Artificial Analysis launched VIALS-AA, its implementation of Handshake AI's VIALS benchmark. It tests whether models can interpret scientific images, such as gels, blots, plasmid maps or flow-cytometry plots, and answer questions about the evidence in them. Each of the 161 tasks is run three times in Stirrup, the firm's open-source reference harness, which gives models code execution and image inspection tools. The score, pass^3, is the share of tasks answered correctly in all three repeats; Artificial Analysis requires all three because a result that changes between runs can't be relied on in scientific work.

As of November 6, 2026, Gemini 4 Carbon (high) led with 64%, answering all three repeats correctly on 103 of 161 tasks, two more than GPT-6 Astra (max). Claude Sonnet 5.5 (max, with fallback) scored 60% and GPT-6.1 Sol (max) 59%, but Sonnet made 3.6× as many tool calls per task. Consistency separated similar models: Kimi K3.1 Pro (max) and GPT-6 Astra (max) each solved about three-quarters of tasks at least once, yet scored 45% and 63% under the all-three rule.

In the launch results, cost per task ranged from $0.02 for GPT-6 Luna (max) to $1.77 for Grok 4.7 (xhigh). GPT-6.1 Sol (max) scored 59% at $0.06 per task, about a thirtieth of the cost of Claude Opus 5.5 (max, with fallback), which scored 61%. Across 19 models and 9,177 rollouts, 38.3% of tool calls viewed images and 47.3% prepared or measured them, for example by cropping, enhancing contrast or measuring pixels.