ParseBench: document parsing for AI agents (LlamaIndex leaderboard) (ParseBench)
We show this table for reference; we do not rank on it.
Scores document parsers and vision-language models on 2,078 human-verified enterprise pages across tables, charts, content faithfulness, semantic formatting and visual grounding, with the listed cost per page.
Overall score on ParseBench — October 25, 2026 snapshot
We mirror the published overall score view for ParseBench. LlamaParse Agentic Plus leads the public snapshot at 88.82%, followed by LlamaParse Agentic (85.90%) and Pulse Ultra 2 (81.60%). We do not use these results to rank models overall.
LlamaParse Agentic Plus
LlamaIndex
LlamaParse · Default · 5.63¢ per page
LlamaParse Agentic
LlamaIndex
LlamaParse · Default · 1.25¢ per page
Pulse Ultra 2
Pulse
Commercial parsing API · Default · 1.50¢ per page
149 configurationsMultimodal & GroundedCurrentDisplay onlyUpdated October 25, 2026 snapshot
Overall score table (149 configurations)
| Rank | Model / configuration | Score | Parameters (B) | Open / closed |
|---|---|---|---|---|
| 1 | LlamaParse Agentic PlusLlamaIndexLlamaParse · Default · 5.63¢ per page | 88.82% | Not reported | Unknown |
| 2 | LlamaParse AgenticLlamaIndexLlamaParse · Default · 1.25¢ per page | 85.90% | Not reported | Unknown |
| 3 | Pulse Ultra 2PulseCommercial parsing API · Default · 1.50¢ per page | 81.60% | Not reported | Unknown |
| 4 | Upstage Document Parse 2 (Enhanced)UpstageProprietary VLM · Enhanced · 3.00¢ per page | 80.51% | Not reported | Unknown |
| 5 | anyformat (Standard)anyformatCommercial parsing API · Standard · 4.30¢ per page | 80.30% | Not reported | Unknown |
| 6 | LlamaParse Cost EffectiveLlamaIndexLlamaParse · Default · 0.38¢ per page | 79.62% | Not reported | Unknown |
| 7 | Claude Opus 5.5AnthropicProprietary VLM · Effort High · 6.13¢ per page | 78.89% | Not reported | Closed |
| 8 | oi-parserCreator not reportedCommercial parsing API · Default · no listed cost | 78.30% | Not reported | Unknown |
| 9 | Claude Fable 5.1AnthropicProprietary VLM · Default · 16.05¢ per page | 78.12% | Not reported | Closed |
| 10 | rakedoc-nanoCloudrakerOpen-weight VLM · Default · no listed cost | 77.23% | Not reported | Unknown |
| 11 | Claude Opus 5.5AnthropicProprietary VLM · Default · 5.79¢ per page | 77.03% | Not reported | Closed |
| 12 | aeddix-alpine-ocr (KDL pipeline)Aeddix LabsOpen-weight VLM · KDL pipeline · no listed cost | 76.77% | Not reported | Unknown |
| 13 | florin-parser-nanoFlorinOpen-weight VLM · Default · no listed cost | 76.69% | Not reported | Unknown |
| 14 | Nutrient DWS (Agentic)NutrientCommercial IDP · Agentic · 1.80¢ per page | 76.43% | Not reported | Unknown |
| 15 | Extend (2.0)ExtendCommercial parsing API · 2.0 · 2.50¢ per page | 75.33% | Not reported | Unknown |
| 16 | Nemotron-3.5-ParseNVIDIAOpen-weight VLM · Default · no listed cost | 75.20% | Not reported | Unknown |
| 17 | LightOnOCR-3-4BLightOnOpen-weight VLM · Default · no listed cost | 75.10% | Not reported | Unknown |
| 18 | Gemini 3 FlashGoogleProprietary VLM · Thinking High · 2.41¢ per page | 75.05% | Not reported | Closed |
| 19 | Databricks AI ParseDatabricksCommercial IDP · Default · 0.35¢ per page | 74.71% | Not reported | Unknown |
| 20 | LightOnOCR-3-0.8BLightOnOpen-weight VLM · Default · no listed cost | 74.60% | Not reported | Unknown |
| 21 | GPT-5.6 SolOpenAIProprietary VLM · Reasoning High · 8.08¢ per page | 74.46% | Not reported | Closed |
| 22 | Infinity-Parser2-ProinflyOpen-weight VLM · Default · no listed cost | 74.28% | Not reported | Unknown |
| 23 | Gemini 3 FlashGoogleProprietary VLM · Thinking Medium · 1.98¢ per page | 74.22% | Not reported | Closed |
| 24 | Claude Opus 5.5AnthropicProprietary VLM · Effort Low · 5.46¢ per page | 73.57% | Not reported | Closed |
| 25 | Gemini 3.5 FlashGoogleProprietary VLM · Thinking High · 4.95¢ per page | 73.36% | Not reported | Closed |
| 26 | Extend Light (1.0)ExtendCommercial parsing API · 1.0 · 0.63¢ per page | 73.26% | Not reported | Unknown |
| 27 | Reducto (Agentic)ReductoCommercial parsing API · Agentic · 4.76¢ per page | 72.97% | Not reported | Unknown |
| 28 | MinerU2.5-Pro-2605-1.2BOpenDataLabOpen-weight VLM · Default · no listed cost | 72.78% | Not reported | Unknown |
| 29 | KDL-Frontier-Parser-nanoKDL AIOpen-weight VLM · Default · no listed cost | 72.38% | Not reported | Unknown |
| 30 | Infinity-Parser2-FlashinflyOpen-weight VLM · Default · no listed cost | 72.18% | Not reported | Unknown |
| 31 | Gemini 3.8 FlashGoogleProprietary VLM · Thinking High · 3.41¢ per page | 71.50% | Not reported | Closed |
| 32 | LightOnOCR-3-1BLightOnOpen-weight VLM · Default · no listed cost | 71.40% | Not reported | Unknown |
| 33 | Gemini 3 FlashGoogleProprietary VLM · Thinking Minimal · 0.65¢ per page | 71.04% | Not reported | Closed |
| 34 | GLM-5.3-FlashZ.AIProprietary VLM · Default · 0.44¢ per page | 70.75% | Not reported | Open |
| 35 | Gemini 3.7 FlashGoogleProprietary VLM · Thinking High · 1.91¢ per page | 70.66% | Not reported | Closed |
| 36 | Chandra-ocr-2DatalabOpen-weight VLM · Default · no listed cost | 70.10% | Not reported | Unknown |
| 37 | Gemini 3.7 FlashGoogleProprietary VLM · Thinking Medium · 1.06¢ per page | 70.09% | Not reported | Closed |
| 38 | Datalab AccurateDatalabCommercial parsing API · Default · 1.00¢ per page | 69.95% | Not reported | Unknown |
| 39 | Gemini 3.8 FlashGoogleProprietary VLM · Thinking Medium · 1.24¢ per page | 69.90% | Not reported | Closed |
| 40 | GPT-6 AstraOpenAIProprietary VLM · Reasoning Low · 9.87¢ per page | 69.88% | Not reported | Closed |
| 41 | Claude Fable 5AnthropicProprietary VLM · Default · 15.60¢ per page | 69.79% | Not reported | Closed |
| 42 | Qwen3.8-27BAlibabaOpen-weight VLM · Thinking · no listed cost | 69.72% | Not reported | Open |
| 43 | GPT-6.1 SolOpenAIProprietary VLM · Reasoning Low · 1.95¢ per page | 69.70% | Not reported | Closed |
| 44 | Datalab BalancedDatalabCommercial parsing API · Default · 0.40¢ per page | 69.51% | Not reported | Unknown |
| 45 | Gemini 3.6 FlashGoogleProprietary VLM · Thinking High · 4.25¢ per page | 69.38% | Not reported | Closed |
| 46 | Gemini 3.8 FlashGoogleProprietary VLM · Thinking Low · 0.71¢ per page | 69.32% | Not reported | Closed |
| 47 | Claude Sonnet 5.5AnthropicProprietary VLM · Default · 2.74¢ per page | 69.26% | Not reported | Closed |
| 48 | Gemini 3.5 FlashGoogleProprietary VLM · Thinking Medium · 4.28¢ per page | 69.20% | Not reported | Closed |
| 49 | Gemini 3.1 ProGoogleProprietary VLM · Default · 8.49¢ per page | 69.14% | Not reported | Closed |
| 50 | GPT-6 SolOpenAIProprietary VLM · Reasoning High · 3.47¢ per page | 68.89% | Not reported | Closed |
| 51 | Gemini 3.7 FlashGoogleProprietary VLM · Thinking Low · 0.70¢ per page | 68.71% | Not reported | Closed |
| 52 | Mistral OCR 4 (Annotation)MistralCommercial IDP · Annotation · 0.50¢ per page | 68.23% | Not reported | Unknown |
| 53 | ReductoReductoCommercial parsing API · Default · 2.38¢ per page | 67.83% | Not reported | Unknown |
| 54 | Datalab FastDatalabCommercial parsing API · Default · 0.40¢ per page | 67.79% | Not reported | Unknown |
| 55 | PaddleOCR-VL-1.6PaddlePaddleOpen-weight VLM · Default · no listed cost | 67.43% | Not reported | Unknown |
| 56 | GPT-6 SolOpenAIProprietary VLM · Reasoning Medium · 2.40¢ per page | 67.38% | Not reported | Closed |
| 57 | GPT-5.6 TerraOpenAIProprietary VLM · Reasoning Medium · 2.71¢ per page | 67.11% | Not reported | Closed |
| 58 | GPT-5.6 LunaOpenAIProprietary VLM · Reasoning XHigh · 1.01¢ per page | 66.89% | Not reported | Closed |
| 59 | GPT-5.6 LunaOpenAIProprietary VLM · Reasoning Max · 1.70¢ per page | 66.85% | Not reported | Closed |
| 60 | GPT-5.6 TerraOpenAIProprietary VLM · Reasoning Low · 2.59¢ per page | 66.74% | Not reported | Closed |
| 61 | GPT-5.5OpenAIProprietary VLM · Reasoning Medium · 13.09¢ per page | 66.63% | Not reported | Closed |
| 62 | Gemini 3.6 FlashGoogleProprietary VLM · Thinking Medium · 3.76¢ per page | 66.06% | Not reported | Closed |
| 63 | GPT-5.6 LunaOpenAIProprietary VLM · Reasoning High · 0.59¢ per page | 66.02% | Not reported | Closed |
| 64 | PaddleOCR-VL-1.5PaddlePaddleOpen-weight VLM · Default · no listed cost | 65.95% | Not reported | Unknown |
| 65 | GPT-6 SolOpenAIProprietary VLM · Reasoning None · 1.86¢ per page | 65.32% | Not reported | Closed |
| 66 | Surya OCR 2DatalabOpen-weight VLM · Default · no listed cost | 64.83% | Not reported | Unknown |
| 67 | DeepSeek V4.1 Flash 1017DeepSeekOpen-weight VLM · Thinking High · no listed cost | 64.56% | Not reported | Open |
| 68 | GPT-6 LunaOpenAIProprietary VLM · Reasoning Max · 0.72¢ per page | 64.26% | Not reported | Closed |
| 69 | Gemini 3.6 FlashGoogleProprietary VLM · Thinking Minimal · 1.48¢ per page | 64.26% | Not reported | Closed |
| 70 | Claude Opus 4.7AnthropicProprietary VLM · Default · 7.14¢ per page | 63.34% | Not reported | Closed |
| 71 | GPT-6 LunaOpenAIProprietary VLM · Reasoning XHigh · 0.29¢ per page | 63.15% | Not reported | Closed |
| 72 | GPT-6 LunaOpenAIProprietary VLM · Reasoning High · 0.21¢ per page | 62.95% | Not reported | Closed |
| 73 | GPT-5.6 TerraOpenAIProprietary VLM · Reasoning None · 2.36¢ per page | 62.83% | Not reported | Closed |
| 74 | GPT-5.5OpenAIProprietary VLM · Reasoning None · 5.93¢ per page | 62.82% | Not reported | Closed |
| 75 | HunyuanOCR 1.5TencentOpen-weight VLM · Default · no listed cost | 62.59% | Not reported | Unknown |
| 76 | Gemma 4 31BGoogleOpen-weight VLM · Default · no listed cost | 62.40% | Not reported | Open |
| 77 | Claude Opus 4.8AnthropicProprietary VLM · Default · 7.30¢ per page | 62.27% | Not reported | Closed |
| 78 | GPT-5.4OpenAIProprietary VLM · Reasoning None · 2.90¢ per page | 62.23% | Not reported | Closed |
| 79 | Gemini 3.5 FlashGoogleProprietary VLM · Thinking Minimal · 1.82¢ per page | 61.99% | Not reported | Closed |
| 80 | Qwen3-VL-8B-InstructAlibabaOpen-weight VLM · Default · no listed cost | 61.97% | Not reported | Unknown |
| 81 | GPT-5.6 LunaOpenAIProprietary VLM · Reasoning Medium · 0.29¢ per page | 61.87% | Not reported | Closed |
| 82 | Reducto (r-1)ReductoCommercial parsing API · r-1 · 1.00¢ per page | 61.68% | Not reported | Unknown |
| 83 | Claude Sonnet 5AnthropicProprietary VLM · Default · 3.18¢ per page | 61.16% | Not reported | Closed |
| 84 | GPT-6 LunaOpenAIProprietary VLM · Reasoning Medium · 0.14¢ per page | 61.10% | Not reported | Closed |
| 85 | GPT-5.6 SolOpenAIProprietary VLM · Reasoning None · 4.06¢ per page | 60.86% | Not reported | Closed |
| 86 | DocAICreator not reportedCommercial parsing API · Default · 1.00¢ per page | 60.74% | Not reported | Unknown |
| 87 | Mistral OCR 4MistralCommercial IDP · Default · 0.40¢ per page | 60.68% | Not reported | Unknown |
| 88 | Qwen3.8-27BAlibabaOpen-weight VLM · Default · no listed cost | 60.30% | Not reported | Open |
| 89 | Gemini 3.1 ProGoogleProprietary VLM · Thinking Low · 2.50¢ per page | 59.87% | Not reported | Closed |
| 90 | Gemini 3.5 Flash-LiteGoogleProprietary VLM · Thinking High · 1.18¢ per page | 59.66% | Not reported | Closed |
| 91 | Azure Document Intelligence (Layout)MicrosoftCommercial IDP · Layout · 1.00¢ per page | 59.64% | Not reported | Unknown |
| 92 | Gemma 4 26B A4BGoogleOpen-weight VLM · Default · no listed cost | 58.50% | Not reported | Open |
| 93 | GPT-5.6 LunaOpenAIProprietary VLM · Reasoning Low · 0.23¢ per page | 58.48% | Not reported | Closed |
| 94 | Gemini 3.1 Flash-LiteGoogleProprietary VLM · Default · 0.29¢ per page | 58.32% | Not reported | Closed |
| 95 | GPT-6 LunaOpenAIProprietary VLM · Reasoning Low · 0.11¢ per page | 58.03% | Not reported | Closed |
| 96 | TeleOCRStarDoc AIOpen-weight VLM · Default · no listed cost | 57.53% | Not reported | Unknown |
| 97 | Gemini 3.1 Flash-LiteGoogleProprietary VLM · Thinking Medium · 0.32¢ per page | 56.83% | Not reported | Closed |
| 98 | Gemini 3.1 Flash-LiteGoogleProprietary VLM · Thinking High · 1.04¢ per page | 56.28% | Not reported | Closed |
| 99 | Claude Haiku 5.5AnthropicProprietary VLM · Default · 0.14¢ per page | 56.10% | Not reported | Closed |
| 100 | Dots.mocrrednoteOpen-weight VLM · Default · no listed cost | 55.79% | Not reported | Unknown |
| 101 | ExtendExtendCommercial parsing API · Default · 2.50¢ per page | 55.75% | Not reported | Unknown |
| 102 | Gemini 3.5 Flash-LiteGoogleProprietary VLM · Default · 0.56¢ per page | 55.50% | Not reported | Closed |
| 103 | DeepSeek V4.1 Flash 1017DeepSeekOpen-weight VLM · No Thinking · no listed cost | 55.20% | Not reported | Open |
| 104 | GPT-5.6 LunaOpenAIProprietary VLM · Reasoning None · 0.22¢ per page | 54.47% | Not reported | Closed |
| 105 | Claude Opus 4.6AnthropicProprietary VLM · Default · 5.78¢ per page | 54.07% | Not reported | Closed |
| 106 | PyMuPDF4LLMArtifexOpen-source local parser · Default · no listed cost | 53.49% | Not reported | Unknown |
| 107 | HPD-ParsingPaddlePaddleOpen-weight VLM · Default · no listed cost | 53.39% | Not reported | Unknown |
| 108 | Claude Haiku 4.5 ThinkingAnthropicProprietary VLM · Thinking · 3.69¢ per page | 53.12% | Not reported | Closed |
| 109 | Gemini 3.5 Flash-LiteGoogleProprietary VLM · Thinking Medium · 0.47¢ per page | 52.22% | Not reported | Closed |
| 110 | Falcon-OCRTIIOpen-weight VLM · Default · no listed cost | 52.07% | Not reported | Unknown |
| 111 | GPT-5 miniOpenAIProprietary VLM · Reasoning Medium · 3.05¢ per page | 51.52% | Not reported | Closed |
| 112 | GPT-6 LunaOpenAIProprietary VLM · Reasoning None · 0.09¢ per page | 51.40% | Not reported | Closed |
| 113 | Docling-modelsDoclingOpen-weight VLM · Default · no listed cost | 50.65% | Not reported | Unknown |
| 114 | Google Cloud Document AIGoogleCommercial IDP · Default · 1.00¢ per page | 50.39% | Not reported | Unknown |
| 115 | Nemotron 3 Nano Omni 30B A3BNVIDIAOpen-weight VLM · Thinking · no listed cost | 48.51% | Not reported | Open |
| 116 | LightOnOCR-2-1BLightOnOpen-weight VLM · Default · no listed cost | 48.00% | Not reported | Unknown |
| 117 | AWS TextractAmazonCommercial IDP · Default · 1.50¢ per page | 47.88% | Not reported | Unknown |
| 118 | GPT-5 miniOpenAIProprietary VLM · Reasoning Minimal · 0.88¢ per page | 46.83% | Not reported | Closed |
| 119 | Qianfan-OCRBaiduOpen-weight VLM · Default · no listed cost | 46.20% | Not reported | Unknown |
| 120 | Unlimited-OCRBaiduOpen-weight VLM · Default · no listed cost | 46.17% | Not reported | Unknown |
| 121 | Jina-OCR-v1Jina AIOpen-weight VLM · Default · no listed cost | 45.93% | Not reported | Unknown |
| 122 | MinerU2.5-2509-1.2BOpenDataLabOpen-weight VLM · Default · no listed cost | 45.90% | Not reported | Unknown |
| 123 | LandingAILandingAICommercial parsing API · Default · 3.00¢ per page | 45.23% | Not reported | Unknown |
| 124 | Claude Haiku 4.5AnthropicProprietary VLM · Disable Thinking · 1.64¢ per page | 45.17% | Not reported | Closed |
| 125 | Qwen3.6-35B-A3BAlibabaOpen-weight VLM · Default · no listed cost | 44.10% | Not reported | Open |
| 126 | GPT-5.4 nanoOpenAIProprietary VLM · Default · 0.25¢ per page | 43.35% | Not reported | Closed |
| 127 | MinerU-Diffusion-V1-0320-2.5BOpenDataLabOpen-weight VLM · Default · no listed cost | 42.63% | Not reported | Unknown |
| 128 | DeepSeek-OCR-2DeepSeekOpen-weight VLM · Default · no listed cost | 41.20% | Not reported | Unknown |
| 129 | PaddleOCR-VLPaddlePaddleOpen-weight VLM · Default · no listed cost | 40.90% | Not reported | Unknown |
| 130 | OvisOCR2Creator not reportedOpen-weight VLM · Default · no listed cost | 40.84% | Not reported | Unknown |
| 131 | Gemma 4 E4BGoogleOpen-weight VLM · Default · no listed cost | 40.50% | Not reported | Open |
| 132 | Amazon Nova 2 LiteAmazonProprietary VLM · Default · 0.70¢ per page | 40.43% | Not reported | Closed |
| 133 | Warp IngestWarpOpen-source local parser · Default · no listed cost | 40.18% | Not reported | Unknown |
| 134 | Granite Vision 4.1 4BIBMOpen-weight VLM · Default · no listed cost | 39.45% | Not reported | Unknown |
| 135 | WeVisDoc-2BTencentOpen-weight VLM · Default · no listed cost | 38.93% | Not reported | Unknown |
| 136 | LiteParse (no OCR)LlamaIndexOpen-source local parser · no OCR · no listed cost | 36.90% | Not reported | Unknown |
| 137 | WeVisDoc-4BTencentOpen-weight VLM · Default · no listed cost | 36.15% | Not reported | Unknown |
| 138 | Qwen3.5-4BAlibabaOpen-weight VLM · Default · no listed cost | 35.40% | Not reported | Unknown |
| 139 | Qwen3.5-9BAlibabaOpen-weight VLM · Default · no listed cost | 31.90% | Not reported | Unknown |
| 140 | FirecrawlFirecrawlCommercial parsing API · Default · 0.90¢ per page | 31.08% | Not reported | Unknown |
| 141 | GLM-OCRZ.AIOpen-weight VLM · Default · no listed cost | 29.60% | Not reported | Unknown |
| 142 | OpenDataLoaderCreator not reportedOpen-source local parser · Default · no listed cost | 29.40% | Not reported | Unknown |
| 143 | Qwen3.5-0.8BAlibabaOpen-weight VLM · Default · no listed cost | 28.40% | Not reported | Unknown |
| 144 | Qwen3.5-2BAlibabaOpen-weight VLM · Default · no listed cost | 27.30% | Not reported | Unknown |
| 145 | pdf-inspectorCreator not reportedOpen-source local parser · Default · no listed cost | 26.59% | Not reported | Unknown |
| 146 | MarkItDownMicrosoftOpen-source local parser · Default · no listed cost | 18.63% | Not reported | Unknown |
| 147 | PyMuPDF (HTML)ArtifexOpen-source local parser · HTML · no listed cost | 16.62% | Not reported | Unknown |
| 148 | PyMuPDF (Text)ArtifexOpen-source local parser · Text · no listed cost | 16.02% | Not reported | Unknown |
| 149 | pypdfCreator not reportedOpen-source local parser · Default · no listed cost | 14.87% | Not reported | Unknown |
ParseBench (Kaggle)
Published Kaggle aggregates only. Model configurations and evaluation dates stay attached to each result. These results do not enter overall or category rankings. Task data, prompts, answers, and per-case outputs are not included. How Kaggle Benchmarks reruns tests.
| Source rank | Source model or configuration | ParseBench score | Source setup | Parameters (B) | Open / closed |
|---|---|---|---|---|---|
| 1 | GPT-6.1 Sol | 80.1% | Task version 4; Evaluated 2026-10-05Original result | Not reported | Unknown |
| 2 | GPT-6 Sol | 80.0% | Task version 4; Evaluated 2026-10-01Original result | Not reported | Unknown |
| 3 | Gemini 3 Flash Preview | 79.3% | Task version 4; Evaluated 2026-04-16Original result | Not reported | Unknown |
| 4 | GPT-6 Astra | 79.1% | Task version 4; Evaluated 2026-09-10Original result | Not reported | Unknown |
| 5 | Gemini 4 Argon | 78.4% | Task version 4; Evaluated 2026-10-09Original result | Not reported | Unknown |
| 6 | GPT-5.6 Terra | 77.5% | Task version 4; Evaluated 2026-08-03Original result | Not reported | Unknown |
| 7 | GPT-5.6 Sol | 77.5% | Task version 4; Evaluated 2026-08-03Original result | Not reported | Unknown |
| 8 | GPT-5.5 | 75.5% | Task version 4; Evaluated 2026-05-23Original result | Not reported | Unknown |
| 9 | GPT-5.4 | 72.9% | Task version 4; Evaluated 2026-04-17Original result | Not reported | Unknown |
| 10 | Gemini 3.7 Flash | 69.5% | Task version 4; Evaluated 2026-08-26Original result | Not reported | Unknown |
| 11 | Grok 4.7 | 69.3% | Task version 4; Evaluated 2026-10-02Original result | Not reported | Unknown |
| 12 | Claude Opus 5 | 69.0% | Task version 4; Evaluated 2026-08-03Original result | Not reported | Unknown |
| 13 | Gemini 3.8 Flash | 67.8% | Task version 4; Evaluated 2026-09-10Original result | Not reported | Unknown |
| 14 | Gemma 4 31B | 66.4% | Task version 4; Evaluated 2026-04-16Original result | Not reported | Unknown |
| 15 | GPT-5.6 Luna | 66.0% | Task version 4; Evaluated 2026-08-03Original result | Not reported | Unknown |
| 16 | Gemini 3.1 Pro Preview | 64.6% | Task version 4; Evaluated 2026-04-17Original result | Not reported | Unknown |
| 17 | Claude Opus 4.7 | 64.0% | Task version 4; Evaluated 2026-05-23Original result | Not reported | Unknown |
| 18 | Claude Opus 4.8 | 64.0% | Task version 4; Evaluated 2026-05-29Original result | Not reported | Unknown |
| 19 | Deepseek V4.1 Flash | 63.8% | Task version 4; Evaluated 2026-10-20Original result | Not reported | Unknown |
| 20 | Gemini 3.1 Flash-Lite Preview | 62.0% | Task version 4; Evaluated 2026-04-16Original result | Not reported | Unknown |
| 21 | Gemma 4 26B A4B | 61.1% | Task version 4; Evaluated 2026-04-16Original result | Not reported | Unknown |
| 22 | Grok 4.6 | 60.7% | Task version 4; Evaluated 2026-09-16Original result | Not reported | Unknown |
| 23 | Claude Haiku 5.5 | 58.5% | Task version 4; Evaluated 2026-10-09Original result | Not reported | Unknown |
| 24 | Claude Sonnet 4.6 | 56.5% | Task version 4; Evaluated 2026-04-16Original result | Not reported | Unknown |
| 25 | Gemini 3.5 Flash-Lite | 55.5% | Task version 4; Evaluated 2026-08-03Original result | Not reported | Unknown |
| 26 | Claude Opus 4.6 | 53.8% | Task version 4; Evaluated 2026-04-17Original result | Not reported | Unknown |
| 27 | GPT-5.4 nano | 49.7% | Task version 4; Evaluated 2026-04-16Original result | Not reported | Unknown |
| 28 | Gemini 2.5 Flash | 49.2% | Task version 4; Evaluated 2026-04-16Original result | Not reported | Unknown |
| 29 | GPT-5.4 mini | 48.3% | Task version 4; Evaluated 2026-04-16Original result | Not reported | Unknown |
| 30 | Gemini 3.5 Flash | 34.0% | Task version 4; Evaluated 2026-05-23Original result | Not reported | Unknown |
| 31 | Gemini 3.6 Flash | 23.7% | Task version 4; Evaluated 2026-08-03Original result | Not reported | Unknown |
| 32 | Grok 4.5 | 21.8% | Task version 4; Evaluated 2026-08-03Original result | Not reported | Unknown |
| 33 | Claude Sonnet 5 | 3.6% | Task version 4; Evaluated 2026-08-03Original result | Not reported | Unknown |
| 34 | DeepSeek V3.2 | 3.3% | Task version 4; Evaluated 2026-04-16Original result | Not reported | Unknown |
How to read this leaderboard
A higher overall score means more of the page structure survived parsing. Read the dimension scores with it: a model can lead on tables and content faithfulness while scoring near zero on visual grounding. Differences of a point or two between reasoning settings of the same model are within normal run variation.
Operator receipt: 149 sourced rows are currently displayable on this page; the leading published row is LlamaParse Agentic Plus at 88.82%.
Honest limit: Results belong to one configuration under LlamaIndex's harness and prompts. Reasoning settings change both score and cost, so a model's best row is not its cheapest. Open-weight and local parsers list no cost because they are self-hosted. LlamaIndex publishes the benchmark and sells LlamaParse, which leads the board; the dataset and evaluation code are public, so the results can be rerun.
How we show ParseBench
LlamaIndex scores document parsers on 2,078 human-verified enterprise pages in five dimensions: tables, charts, content faithfulness, semantic formatting and visual grounding. The overall score is the mean of the five. We mirror the leaderboard CSV the public site loads: 149 configurations across 113 parsers and models, with cost in cents per page where a list price exists.
Vision-language models appear once per reasoning setting, because the setting changes both score and cost. Each catalog model's page shows its best configuration with the setting in its note; parsing APIs, OCR specialists and local parsers keep their source names.
Every row is one configuration under LlamaIndex's harness, and LlamaIndex sells LlamaParse, which leads the board. The dataset and evaluation code are public. The table is display only and does not enter model rankings. Kaggle Benchmarks reruns the task through its own harness with different numbers; that board is the separate table below.
Snapshot
The published ParseBench snapshot places LlamaParse Agentic Plus first at 88.82%. The third row is 7.22 points behind. The broader top-10 range is 11.59 points, so the table still separates the published systems.
149 configurations have been evaluated on ParseBench. The benchmark falls in the Multimodal & Grounded category. ParseBench is currently displayed for reference but excluded from the scoring formula, so it does not directly affect overall rankings.
About ParseBench
Year
2026
Tasks
2,078 human-verified enterprise pages from insurance, finance and government documents
Format
Deterministic test rules per page; five dimension scores and their mean, in percent
Difficulty
No method is strong on all five dimensions; charts and visual grounding separate the field
ParseBench judges a parser on whether its output keeps the structure an agent needs rather than on text similarity. Each page carries deterministic test rules in five dimensions: tables (structure, merged cells, hierarchical headers), charts (data points, series, axes), content faithfulness (omissions and hallucinations), semantic formatting (strikethrough, superscripts, emphasis, heading levels) and visual grounding (element localization and classification). The overall score is the mean of the five dimension scores. The leaderboard lists every configuration LlamaIndex ran, from commercial parsing APIs and cloud document services to proprietary and open-weight vision-language models at each reasoning setting and local open-source parsers, with cost in cents per page where a list price exists. We mirror the leaderboard CSV from the ParseBench repository, which the public site loads at runtime, and store each catalog model's best configuration on its row.
Freshness and provenance
Version
ParseBench leaderboard (LlamaIndex repository CSV)
Refresh cadence
Repository leaderboard updates
Staleness state
Current
Question availability
The 2,078-page dataset and evaluation code are public under Apache 2.0
BenchLM uses freshness metadata to decide whether a benchmark should still be treated as a strong differentiator, a benchmark to watch, or a display-only reference. For the full scoring policy, see the BenchLM methodology page.
Questions
What does ParseBench measure?
ParseBench measures whether a parser's output keeps what an agent needs from a page: table structure, chart values, the full text without omissions or inventions, meaningful formatting and the position of each element. Every page carries deterministic test rules, and the overall score is the mean of the five dimension scores.
How does this table differ from the Kaggle rerun below it?
This table mirrors LlamaIndex's own leaderboard: 148 configurations including commercial parsing APIs, open-weight OCR models and local parsers. Kaggle Benchmarks reruns the same task on 33 models through its own harness and publishes different numbers for the same models, so its board is shown as a separate table on this page and never merged.
Why is ParseBench display only?
Each row is one configuration under LlamaIndex's harness, reasoning settings change both score and cost, and the publisher sells the leading system. The board is public and reproducible, which is why we show every row, but the results compare those exact runs and do not enter BenchLM's model rankings.
Know when it’s worth switching models
The model to choose, the cheaper alternative, and the release we would wait on.
Read a sample issueJoin 6,100+ subscribers.
One email each week. Unsubscribe anytime.