SimpleQA Verified (Epoch anti-abstention)
General QA · 2026-08-27
Epoch's August 27, 2026 protocol for the 1,000-question SimpleQA Verified dataset adds an explicit guessing instruction that discourages abstention. Scores are percentage correct, rather than Google's F1 measure. Kept separate from bare-question evaluations.
Top models (higher is better)
| Model | Score |
|---|---|
| GPT-6 Astra | 75.6 |
| Claude Fable 5.1 | 70.8 |
| Gemini 3.8 Flash | 69.7 |
| Gemini 3.7 Flash | 69.2 |
| Gemini 3 Flash Preview | 66.8 |
| Gemini 3.5 Flash | 66.2 |
| Gemini 3.6 Flash | 66.2 |
| GPT-5.5 | 63.0 |
| Muse Spark 1.2 | 60.3 |
| Muse Spark 1.1 | 57.8 |
| Qwen3.7-Max | 55.8 |
| Opus 4.8 | 53.0 |
| DeepSeek V4 Pro 0813 | 52.9 |
| Qwen3.6-Max-Preview | 52.0 |
| Opus 4.7 | 51.7 |
| Kimi K3 | 50.6 |
| GPT-5 | 50.1 |
| o3 | 49.4 |
| Grok 4.6 | 49.3 |
| Qwen3-Max (2025-09-23) | 48.7 |
| Grok 4.5 | 48.3 |
| GPT-5.1 | 48.0 |
| Qwen3.8 Max (0902) | 47.3 |
| Opus 4.6 | 47.0 |
| DeepSeek-V4-Pro | 47.0 |
| GPT-5.4 Pro | 46.3 |
| Qwen3.8-Max | 45.8 |
| Opus 4.5 | 45.7 |
| GPT-5.4 | 45.1 |
| Qwen3.6 Plus (2026-04-02) | 44.1 |
| O1 | 41.1 |
| GLM-5.3 | 41.0 |
| Qwen3 235B A22B Thinking 2507 | 40.4 |
| Inkling | 40.3 |
| GPT-5.2 | 37.1 |
| Kimi K2.7 Code | 36.5 |
| Kimi K2.5 | 34.3 |
| GLM-5.2 | 34.2 |
| GLM-5.1 | 34.0 |
| Sonnet 5 | 33.7 |