DTBench
General QA · 2024-11-15
DTBench (Decision Theory Benchmark) tests reasoning about the decision theory of Newcomb-like problems, such as interactions between an agent and near-copies of itself, using handcrafted multiple-choice questions validated by two decision theorists. Scores are accuracy on the 375 non-trivia capability questions; trivia and the 130 attitude questions, which have no agreed answer, are not scored. The maintainers treat 40% as the random-guessing level because answer-option counts vary.
Top models (higher is better)
| Model | Score |
|---|---|
| Claude Opus 5.5 | 98.9 |
| Claude Fable 5 | 98.4 |
| Claude Opus 5 | 97.9 |
| Claude Fable 5.1 | 97.6 |
| Gemini 4 Argon | 97.6 |
| Claude Sonnet 5.5 | 97.3 |
| GPT-6 Astra | 97.3 |
| GPT-6 Sol | 97.3 |
| Grok 4.6 | 97.3 |
| Gemini 3.1 Pro Preview | 97.1 |
| Gemini 3.7 Flash | 96.8 |
| GPT-6.1 Sol | 96.5 |
| Muse Spark 1.3 | 96.5 |
| GPT-5.5 Pro | 96.0 |
| GPT-5.5 | 96.0 |
| Grok 4.7 | 96.0 |
| GPT-5.6 Sol | 95.7 |
| Gemini 3.8 Flash | 95.7 |
| Gemini 3.6 Flash | 95.5 |
| Muse Wonder | 95.2 |
| Claude Haiku 5.5 | 95.0 |
| Opus 4.8 | 94.9 |
| Gemini 3.5 Flash | 94.7 |
| Opus 4.7 | 94.7 |
| GPT-5.4 | 94.4 |
| Muse Spark 1.1 | 94.4 |
| DeepSeek V4 Pro 0813 | 93.9 |
| GPT-5.6 Terra | 93.3 |
| Qwen3.8 2.4T A95B | 93.2 |
| Sonnet 5 | 92.5 |
| Qwen3.7-Max (source unspecified) | 92.3 |
| Qwen3.8-Max | 92.0 |
| Opus 4.6 | 91.2 |
| Kimi K3 | 91.2 |
| DeepSeek-V4-Flash-0731 | 90.9 |
| GPT-5.2 | 90.9 |
| Kimi K2.6 | 90.9 |
| DeepSeek-V4-Pro | 90.7 |
| GPT-5 | 90.7 |
| GPT-5.1 | 90.1 |