Handshake AI’s VIALS benchmark gets an independent leaderboard, and the results reveal a 99x cost gap between top models

Reading a gel blot or interpreting a microscopy image requires the kind of domain knowledge that takes researchers years to develop. VIALS-AA, launched by Artificial Analysis on November 6, 2026, is a direct attempt to close that measurement gap.

The leaderboard is an independent implementation of Handshake AI’s VIALS benchmark, built specifically to stress-test vision-language models on life sciences and biotech imagery: gel blots, molecular structures, microscopy slides, and similar artifacts that show up constantly in real research workflows but almost never in standard AI evaluations.

Who’s leading and what it costs them

Across 161 visual question-answering tasks, Gemini 4 Carbon (high) currently holds the top position with a score of 64%, measured as the proportion of tasks answered correctly across three repeated runs. GPT-6 Astra (max) sits just behind at 63%, with Claude Opus 5.5 (max with fallback) at 61%.

GPT-6 Luna (max) completes the full task set for $0.02. Grok 4.7 (xhigh) runs the same gauntlet for $1.77. That is a price difference of nearly 99 times for models operating in broadly the same performance tier.

The comparison between GPT-6.1 Sol (max) and Claude Opus 5.5 makes the tradeoff even starker. Sol scores 59% at approximately $0.06 per task. Opus 5.5 scores 61% for approximately $1.64 per task, a 29x cost difference for a 2-percentage-point edge in accuracy.

The open-weights gap and the safety refusal problem

Kimi K3.1 Pro, the benchmark’s leading open-weights model, scored 45%, approximately 19 percentage points behind the top-performing proprietary systems.

Separate from the accuracy scores, fourteen tasks across the benchmark triggered safety refusals from various models. The affected tasks were not flagged for being dangerous; they were flagged because models misread scientific imagery as potentially sensitive content and declined to engage.

The evaluation itself runs through the Stirrup harness, which allows models to execute code and inspect images programmatically rather than just receiving them as static inputs. Performance is measured using a pass^3 metric, meaning a model must answer correctly on all three attempts at a given task to get credit.

Why this benchmark exists now

Handshake AI published the original VIALS dataset in August 2026, with a revision following on September 19. The dataset was made publicly available with the explicit goal of building evaluation infrastructure for real-world scientific image interpretation.

Artificial Analysis built VIALS-AA as an independent implementation, running the benchmark continuously across a broad set of frontier models rather than as a one-time published result.