Artificial Analysis's independent GDP.pdf run: 100 professional PDF tasks, All-pass over 1,275 criteria, LiteParse text plus page images, GPT-5.6 Luna Medium judge.
active
Recorded reasons:
| Category | long-context |
|---|---|
| Subcategory | independent re-score of Surge AI GDP.pdf (professional PDF reasoning) |
| Page status | active |
| Metric | All-pass rate |
| Direction | higher_is_better |
| Unit | % |
| Dataset size | 100 |
| Publisher | Artificial Analysis |
GDP.pdf-AA is Artificial Analysis's implementation of Surge AI's GDP.pdf. Each item is one professional question plus one original PDF. The model must ground the answer in that file, including tables, charts, footnotes, legends, and later amendments. Coverage is 100 tasks in ten domains. AA delivers extracted page text to every model and page images only to models that accept image input. Answers are single-turn free text, with no tools and no browsing.
One PDF per task. AA prepares the file with LiteParse (OCR where needed). Every model gets full extracted text. Image-capable models also get ordered page images (150 DPI, reduced as far as 72 DPI under payload limits; composites of two or four pages when an endpoint caps image count). Five independent attempts. GPT-5.6 Luna Medium grades each atomic criterion without the source PDF or the contestant name.
Each row was checked against its source by a reviewer.
| Model | Score | Evidence date | Source kind | Link |
|---|---|---|---|---|
| GPT-6 Astra (max) | 33.0% | 2026-09-04 published | independent_evaluator | source |
| GLM-5.3 (max) | 12.0% | 2026-09-04 published | independent_evaluator | source |
No model card in ModelSpec reports this benchmark yet.