ParseBench

Not yet imported

Document parsing evaluated across tables, charts, content faithfulness, semantic formatting, and visual grounding.

2,078 pagesApache-2.0version upstream

Listed, not imported — scores come from LlamaIndex's official leaderboard.

View official leaderboard ↗

From the authors

Leaderboard

Sorted by overall score, then lower cost.

upstream ↗ @ cb4c20d · captured 2026-10-20

Showing overall score; higher is better. 148 systems.

ParseBench costs are converted from cents to USD. The leaderboard CSV does not report sample counts; dimension counts are left unspecified.

Overall score leaderboard, source: upstream
#	system	overall score ↓	tables—	charts—	content faithfulness—	semantic formatting—	visual grounding—	¢ / page
1	L LlamaParse Agentic PlusLlamaIndex · llamaparse	90.2%	93.4%	94.2%	92.3%	87.1%	84.1%	5.63
2	L LlamaParse AgenticLlamaIndex · llamaparse	87.0%	88.9%	88.7%	91.8%	81.4%	84.3%	1.25
3	P Pulse Ultra 2Pulse · commercial - startup apis	81.6%	90.3%	89.7%	87.6%	74.0%	66.3%	1.50
4	L LlamaParse Cost EffectiveLlamaIndex · llamaparse	80.6%	84.2%	77.9%	89.9%	67.3%	83.8%	0.38
5	A anyformat (Standard)anyformat · commercial - startup apis	80.3%	86.5%	85.6%	87.3%	79.0%	63.1%	4.30
6	A Anthropic Opus 5.5 (Effort High)Anthropic · vlm - proprietary	79.8%	94.3%	70.9%	91.7%	77.0%	65.3%	6.13
7	A Anthropic Fable 5.1Anthropic · vlm - proprietary	78.9%	91.5%	67.1%	91.2%	76.5%	68.3%	16.05

dimension score0100Bold marks the best score in each dimension.

Quality against costOverall score by cost per page. Systems without positive reported cost are omitted.

llamaparse / vlm - proprietary / commercial - idp / commercial - startup apis / vlm - open weight / open source - local

Head to head

First system

Second system

LLlamaParse Agentic Plus
LLlamaParse Agentic

What it measures

ParseBench measures how faithfully document parsers preserve information for AI agents. Its five dimensions expose errors that a single text-similarity score can miss.