Benchmarks / Factuality
MASK
Honesty-under-pressure benchmark: does the model contradict its own beliefs?
maintained by Scale (Scale Labs) · 67 observations on record
Versions
1 version
| Version | Tasks |
|---|---|
| MASK current mask | — |
Current leaderboard
MASK · Accuracy · top 25 of 67 systems
| # | System | Accuracy % |
|---|---|---|
| 1 | Claude Opus 4.6no reasoning Anthropic | 96.3% 95.9–96.7 |
| 2 | Claude Sonnet 4.5 20250929thinking Anthropic | 96.1% 95.6–96.7 |
| 3 | Claude Sonnet 4thinking Anthropic | 95.3% 93.0–97.6 |
| 4 | Claude Opus 4.1 20250805thinking Anthropic | 94.2% 92.4–96.0 |
| 5 | Claude Opus 4.5 20251101thinking Anthropic | 92.5% 91.3–93.8 |
| 6 | GPT OSS 120B OpenAI | 92.0% 91.1–92.9 |
| 7 | GPT 5.4 Pro (2026-03-05) OpenAI | 91.7% 89.7–93.7 |
| 8 | GPT 5.4 (2026-03-05)max effort OpenAI | 89.7% 88.9–90.4 |
| 9 | Claude Sonnet 4 Anthropic | 89.3% 87.3–91.3 |
| 10 | Claude Opus 4thinking Anthropic | 87.9% 84.1–91.6 |
| 11 | Claude Opus 4.1 20250805 Anthropic | 87.4% 85.7–89.1 |
| 12 | Claude Opus 4.5 20251101 Anthropic | 87.1% 86.1–88.2 |
| 13 | GPT 5.2 (2025-12-11) OpenAI | 86.7% 84.9–88.4 |
| 14 | GPT OSS 20B OpenAI | 86.5% 84.4–88.5 |
| 15 | Claude Sonnet 4.5 20250929 Anthropic | 86.4% 84.9–87.9 |
| 16 | GPT 5.1 (2025-11-13)thinking OpenAI | 86.3% 85.6–87.1 |
| 17 | GPT 5 Pro (2025-10-06) OpenAI | 86.0% 84.4–87.6 |
| 18 | Claude Opus 4.6max effort Anthropic | 85.4% 84.9–85.9 |
| 19 | o3 (high) (April 2025) OpenAI | 84.5% 82.1–86.8 |
| 20 | o3 (medium) (April 2025) OpenAI | 82.6% 79.8–85.4 |
| 21 | GPT 5 Mini (2025-08-07) OpenAI | 82.6% 80.3–84.9 |
| 22 | o3 Pro (high) (June 2025) OpenAI | 82.5% 80.9–84.1 |
| 23 | Claude 3.7 Sonnet (February 2025)thinking Anthropic | 82.1% 80.9–83.4 |
| 24 | Claude Opus 4 Anthropic | 80.3% 79.7–80.9 |
| 25 | GPT 5 (2025-08-07) OpenAI | 79.3% 74.0–84.6 |
One row per evaluated system — reasoning-effort variants rank separately.