Artificial Analysis releases VIALS-AA and analyzes frontier models

On 2026-11-07, Artificial Analysis announced on X that it has launched VIALS-AA, a leaderboard for Handshake AI's VIALS benchmark covering visual reasoning in life sciences and biotech research. Image-like visual artifacts are central to how scientists analyze experiments and make research decisions, but correctly interpreting them remains unsolved for AI systems.

The same day, Artificial Analysis posted several VIALS-AA data points. On cost efficiency, GPT-6.1 Sol scores 59% at roughly $0.056 per task and GPT-6.1 Terra scores 60% at about $0.073 per task; in comparison, GPT-6 Astra reaches 63% at roughly 5x that cost per task, while Gemini 4 Carbon reaches 64% at roughly 23x. The account also highlighted how visual reasoning differs by internal setup: Claude Opus 5.5 (max with fallback) averages about 32 tool calls per task, versus roughly 8 for GPT-6 Astra, even though their VIALS-AA scores are similar (61% and 63%, respectively). The takeaway, in Artificial Analysis's framing, is that similar leaderboard scores can come from very different visual-reasoning approaches.

In the reporting available here, Artificial Analysis also described VIALS consistency for Kimi K3.1 Pro versus GPT-6. Across three attempts, Kimi K3.1 Pro answered 119 of 161 tasks correctly at least once, compared with 118 of 161 for GPT-6. The account stressed that solving a task once does not mean a model will answer it consistently, and that consistency is a key part of the advantage for frontier models on VIALS. No later updates are included in these sources, so the current public state is the launch plus these measurements.