Document parsing
ParseBench
Deterministic PDF parsing across tables, charts, content faithfulness, semantic formatting, and visual grounding.
As of 13 August 2026. Published benchmark. Official ParseBench leaderboard observed 13 August 2026. Five dimensions. No overall column.
ParseBench is a deterministic PDF parse audit: tables, charts, content faithfulness, semantic formatting, and visual grounding. Those metrics do not mean the same thing. An unweighted overall is omitted because it would rank a system that fails grounding ahead of a system that an audit needs.
Cents per page is a parser-layer list cost, not full cost per accepted research output. Native PDF paths in frontier VLMs are a provider document harness, not a raw-model cell. Parser vendors may optimize against a public leaderboard; treat current ranks as dated observations.
- Scale
- 2,078 unique pages, 1,211 documents, 169,011 deterministic rules
- Provenance
- LlamaIndex ParseBench repository and arXiv 2604.08538. Apache-2.0 as observed.
- Contamination
- Moderate. Public since 2026. Parser vendors may optimize against it.
- Cannot establish
- Downstream diligence correctness, a document-product moat, or operating economics.
Metrics, kept separate
- GTRM (tables)
- ChartDataPointMatch
- Content Faithfulness Score
- Semantic Formatting Score
- Element Pass Rate
Third-party leaderboard · not an Antidote run · not a product score
Official ParseBench leaderboard, observed 13 August 2026. Five dimensions kept separate. Gemini 3 Flash is Thinking High. GPT-5.6 Sol and Fable 5 are Reasoning None / default API settings in the source.
| Dimension | Gemini 3 Flash | GPT-5.6 Sol | Claude Fable 5 | Claude Opus 4.8 |
|---|---|---|---|---|
| Tables | 91.5% | 89.34% | 89.79% | 89.65% |
| Charts | 64.79% | 59.57% | 52.21% | 49.75% |
| Faithfulness | 90.87% | 86.17% | 90.02% | 89.02% |
| Formatting | 68.31% | 50.42% | 72.62% | 71.38% |
| Grounding | 59.77% | 25.09% | 49.24% | 18.69% |
Third-party leaderboard · not an Antidote run · not a product score
Same ParseBench leaderboard. LlamaParse Agentic, Pulse Ultra 2, and Reducto Agentic against Gemini 3 Flash Thinking High. The source Overall column is omitted.
| Dimension | LlamaParse Agentic | Gemini 3 Flash | Pulse Ultra 2 | Reducto Agentic |
|---|---|---|---|---|
| Tables | 90.74% | 91.5% | 75.45% | 80.42% |
| Charts | 78.11% | 64.79% | 90.82% | 73.4% |
| Faithfulness | 89.68% | 90.87% | 79.49% | 86.37% |
| Formatting | 85.24% | 68.31% | 73.05% | 57.6% |
| Grounding | 80.62% | 59.77% | 66.56% | 67.07% |
ParseBench official leaderboard, observed 13 August 2026. Horizontal axis is list cents per page. Vertical axis is the Tables metric (GTRM), not an overall score.
Third-party leaderboard · not an Antidote run · not a product score
What the figures bound
On the observed leaderboard, Gemini 3 Flash (Thinking High) reaches 91.50% on tables and 59.77% on grounding at 2.41 cents per page. GPT-5.6 Sol is 89.34% tables and 25.09% grounding at 5.87 cents. Claude Fable 5 is 89.79% tables at 15.60 cents. LlamaParse Agentic holds 80.62% grounding at 1.25 cents. Pulse Ultra 2 leads charts (90.82%) at 15 cents per page.
A higher list price does not buy grounding. A specialist parser can beat a frontier VLM on the dimension auditability needs. None of this is a diligence-memo score, a Hebbia or AlphaSense result, or operating economics for a research workflow.
What this page does not conclude
- Downstream diligence or memo quality
- A document-product moat
- Full cost per accepted output
- An overall ParseBench winner