# CritPt v2 Contamination Audit

> CritPt v2 scores 8/10 (Tier A, Low Risk) on BenchAudit as of November 2026, reflecting public prompts with private held-out answers and isolated offline third-party grading.

- **Canonical URL:** https://benchaudit.org/benchmarks/critpt-v2/
- **Benchmark ID:** `critpt-v2`
- **Hub:** Artificial Analysis / CritPt
- **Category:** Science
- **Domains:** Science, Coding
- **Skills:** Reasoning, Coding, Tool Use, Agentic
- **BenchAudit Score:** 8/10 (Tier A - Low Risk)
- **Last Audited:** 2026-11-17 (Ref 2026-11)

## Factor Scores

| Factor | Score (/10) | Weight |
| --- | ---: | ---: |
| Privacy | 5.5 | 40% |
| Originality | 10 | 25% |
| Freshness | 9 | 20% |
| Grading | 10 | 15% |

## Primary-Source Evidence

### Privacy
> "To ensure data integrity and prevent leakage, our grading server only accepts complete batches containing responses to all 70 problems. Any submission with missing or unmatched problem IDs will be automatically rejected."
- Source: https://github.com/CritPt-Benchmark/CritPt

### Originality
> "CritPt consists of 71 composite research challenges designed to simulate full-scale research projects at the entry level, which are also decomposed to 190 simpler checkpoint tasks for more fine-grained insights. All problems are newly created by 50+ active physics researchers based on their own research."
- Source: https://arxiv.org/abs/2509.26574

### Freshness
> "Version 5.0 October 2026 to November 2026 Added HLE-Diamond (10%) to the General category Added Terminal-Bench-Science (15%) to the Science category Added FrontierCode-AA 2.0 (15%) to the Coding category Reduced the GDP.pdf weight from 10% to 5% Replaced CritPt with CritPt v2 (Hard Subset), keeping its 10% weight"
- Source: https://artificialanalysis.ai/methodology/intelligence-benchmarking

### Grading
> "CritPt v2 is the revised version of CritPt, a benchmark designed to test LLMs on research-level physics reasoning tasks, featuring 70 composite research challenges. Models solve problems with a code-execution tool in an offline Linux sandbox."
- Source: https://artificialanalysis.ai/evaluations/critpt-v2

## Top Frontier Models on CritPt v2

| Rank | Model | Creator | Percentile |
| ---: | --- | --- | ---: |
| #1 | [Claude Fable 5.2](https://benchaudit.org/models/claude-fable-5-2__max/) | Anthropic | 91.1p |
| #2 | [Muse Wonder Full](https://benchaudit.org/models/muse-wonder-full__max/) | Meta | 88.7p |
| #3 | [Claude Opus 5.5](https://benchaudit.org/models/claude-opus-5-5__max/) | Anthropic | 84.0p |
| #3 | [GPT-6 Astra](https://benchaudit.org/models/gpt-6-astra__max/) | OpenAI | 84.0p |
| #3 | [GPT-6.1 Sol](https://benchaudit.org/models/gpt-6-1-sol__max/) | OpenAI | 84.0p |
| #6 | [Claude Sonnet 5.5](https://benchaudit.org/models/claude-sonnet-5-5__max/) | Anthropic | 79.3p |
| #7 | [GPT-6 Sol](https://benchaudit.org/models/gpt-6-sol__max/) | OpenAI | 76.9p |
| #8 | [Claude Fable 5.1](https://benchaudit.org/models/claude-fable-5-1__max/) | Anthropic | 74.6p |
| #9 | [Muse Wonder](https://benchaudit.org/models/muse-wonder__max/) | Meta | 72.2p |
| #10 | [GPT-6.1 Terra](https://benchaudit.org/models/gpt-6-1-terra__max/) | OpenAI | 69.8p |
