Gemini 4 Carbon tops VIALS-AA scientific image benchmark

Artificial Analysis published results from VIALS-AA, its implementation of Handshake AI's VIALS benchmark. The benchmark tests whether language models can interpret scientific images and answer questions about the evidence in them. Each of the 161 tasks runs three times in Stirrup, the firm's open-source harness, which gives models code execution and image inspection tools. A task counts as solved only if all three runs are correct, because, in the firm's words, a result that changes between runs cannot be relied on in scientific work.

Gemini 4 Carbon (high) leads at 64%, solving 103 tasks on every repeat. That is two more than GPT-6 Astra (max), which scores 63%. The consistency requirement matters. Kimi K3.1 Pro (max) and GPT-6 Astra each solve about three-quarters of tasks at least once, but Kimi scores 45% under the stricter metric.

Cost per task ranges from $0.02 for GPT-6 Luna (max) to $1.77 for Grok 4.7 (xhigh). GPT-6.1 Sol (max) scores 59% at $0.06 per task, about a thirtieth of the cost of Claude Opus 5.5 (max, with fallback), which scores 61%.

More tool use did not reliably improve accuracy. Claude Sonnet 5.5 (60%) made 3.6× as many tool calls per task as GPT-6.1 Sol. On a gel electrophoresis task, GPT-6 Astra answered correctly in six turns. Grok 4.7 used 27 turns and 20 image views but anchored on the printed label instead of the protein spot and gave the wrong answer.

The writeup also tested GLM-5.3 (max), which lacks native vision, working only through code. Its score is not given in the excerpt.