GDP.pdf
Professional Work · 2026-04-14
Surge AI's GDP.pdf evaluates professional document reasoning on 100 held-out PDF tasks with five runs per task. The Inspect AI harness uses no tools and Gemini 3.5 Flash as judge. Headline accuracy averages the all-rubric-criteria-pass indicator over runs and tasks. Distinct from OpenAI's GDPval benchmark.
Top models (higher is better)
| Model | Score |
|---|---|
| GPT-6.1 Sol | 35.0 |
| GPT-6 Astra | 34.2 |
| Muse Wonder Full | 33.2 |
| Claude Sonnet 5.5 | 31.4 |
| Muse Wonder | 30.8 |
| GPT-5.6 Sol | 30.7 |
| Claude Opus 5.5 | 30.6 |
| Claude Fable 5 | 29.8 |
| Gemini 4 Argon | 28.4 |
| Claude Fable 5.1 | 27.6 |
| Muse Spark 1.3 | 27.6 |
| GPT-6 Sol | 26.4 |
| GPT-5.5 | 26.0 |
| GPT-5.6 Terra | 24.7 |
| Opus 4.8 | 24.0 |
| Claude Opus 5 | 24.0 |
| Gemini 3.7 Flash | 23.8 |
| Gemini 3.8 Flash | 23.2 |
| GPT-6 Luna | 23.0 |
| Grok 4.7 | 22.8 |
| GPT-5.6 Luna | 22.7 |
| Opus 4.7 | 21.0 |
| DeepSeek V4.1 Flash | 19.8 |
| Kimi K3 | 19.0 |
| Sonnet 4.6 | 18.0 |
| Grok 4.6 | 17.2 |
| Gemini 3.1 Pro Preview | 17.0 |
| DeepSeek V4 Flash Vision Exp | 16.2 |
| Muse Spark 1.2 | 16.0 |
| Gemini 3.6 Flash | 14.0 |
| GLM-5.3 Flash | 14.0 |
| Grok 4.5 | 14.0 |
| Kimi K2.6 | 12.0 |
| Muse Glimmer 30B | 11.8 |
| Gemini 3.5 Flash-Lite | 10.0 |
| Gemini 3 Flash Preview | 10.0 |
| Grok 4.3 | 8.0 |
| Mistral Large 3 675B Instruct 2512 | 2.0 |
| Nemotron 3 Nano Omni 30B A3B | 2.0 |