# VIALS-AA Contamination Audit

> VIALS-AA scores 5.4/10 (Tier B, Moderate Risk) on BenchAudit as of November 2026, reflecting publicly downloadable plaintext test files and isolated offline third-party grading.

- **Canonical URL:** https://benchaudit.org/benchmarks/vials-aa/
- **Benchmark ID:** `vials-aa`
- **Hub:** Artificial Analysis / Handshake AI
- **Category:** Science
- **Domains:** Science
- **Skills:** Multimodal, Reasoning, Agentic, Tool Use
- **BenchAudit Score:** 5.4/10 (Tier B - Moderate Risk)
- **Last Audited:** 2026-11-17 (Ref 2026-11)

## Factor Scores

| Factor | Score (/10) | Weight |
| --- | ---: | ---: |
| Privacy | 0 | 40% |
| Originality | 10 | 25% |
| Freshness | 7 | 20% |
| Grading | 10 | 15% |

## Primary-Source Evidence

### Privacy
> "We use the same 161 tasks as Handshake AI's upstream VIALS implementation, covering 174 source images and 8 artifact domains. Each model makes 3 independent repeats per task."
- Source: https://artificialanalysis.ai/methodology/intelligence-benchmarking

### Originality
> "We introduce VIALS, a visual question-answering benchmark with 161 such interpretation tasks, spanning the types of artifacts examined throughout experimental workflows in the biotech industry (rather than polished figures from publications and textbooks)."
- Source: https://arxiv.org/abs/2608.21357

### Freshness
> "Models must be rerun on these tasks. Two prompts were updated to reduce remaining ambiguity, one reference answer and one image were updated. Released 2026-10-30 as the `v2.0.0` tag."
- Source: https://huggingface.co/datasets/handshake-ai-research/VIALS/blob/main/CHANGELOG.md

### Grading
> "Models receive the question and image file paths in the prompt, with the original image files available in a sandbox for code execution with Python and no network access."
- Source: https://artificialanalysis.ai/evaluations/vials-aa

## Top Frontier Models on VIALS-AA

| Rank | Model | Creator | Percentile |
| ---: | --- | --- | ---: |
| #1 | [Claude Fable 5.2](https://benchaudit.org/models/claude-fable-5-2__max/) | Anthropic | 100.0p |
| #2 | [GPT-6.2 Astra](https://benchaudit.org/models/gpt-6-2-astra__max/) | OpenAI | 95.5p |
| #3 | [Gemini 4 Carbon](https://benchaudit.org/models/gemini-4-carbon__high/) | Google | 90.9p |
| #4 | [GPT-6 Astra](https://benchaudit.org/models/gpt-6-astra__max/) | OpenAI | 84.1p |
| #4 | [GPT-6.2 Sol](https://benchaudit.org/models/gpt-6-2-sol__max/) | OpenAI | 84.1p |
| #6 | [Claude Opus 5.5](https://benchaudit.org/models/claude-opus-5-5__max/) | Anthropic | 77.3p |
| #7 | [GPT-6.1 Terra](https://benchaudit.org/models/gpt-6-1-terra__max/) | OpenAI | 72.7p |
| #8 | [Claude Sonnet 5.5](https://benchaudit.org/models/claude-sonnet-5-5__max/) | Anthropic | 65.9p |
| #8 | [Gemini 4 Argon](https://benchaudit.org/models/gemini-4-argon__high/) | Google | 65.9p |
| #10 | [GPT-6.1 Sol](https://benchaudit.org/models/gpt-6-1-sol__max/) | OpenAI | 59.1p |
