New: Agent Index for knowledge work is live.Explore the index ↗
Compare models

Best LLMs for professional document reasoning

GDP.pdf

GDP.pdf tests reasoning over long professional documents across ten domains. The All-pass score measures attempts satisfying every task criterion. These results use the independent five-attempt evaluation with LiteParse text and page images for models supporting images; other implementations remain separate.

Updated: Data as of:
Widely usedNot saturated
Official benchmark
GDP.pdf results. Each row identifies its benchmark version and evaluation track. Scores from different versions are not directly comparable. Best effort from the latest stored report date per model and version.
Model / creatorVersionReported score
GPT-6 AstraOpenAI · Proprietary100 tasks / All-pass / five attempts32.2%
GPT-6.1 SolOpenAI · Proprietary100 tasks / All-pass / five attempts32%
Kimi K3.1 ProMoonshot AI · Open source / open weights100 tasks / All-pass / five attempts31.4%
Muse Wonder 1.0Meta · Proprietary100 tasks / All-pass / five attempts30.6%
Gemini 4 CarbonGoogle · Proprietary100 tasks / All-pass / five attempts29.6%
GPT-6.1 TerraOpenAI · Proprietary100 tasks / All-pass / five attempts28.6%
GPT-5.6 SolOpenAI · Proprietary100 tasks / All-pass / five attempts27.8%
Muse Spark 1.3Meta · Proprietary100 tasks / All-pass / five attempts26.6%
Claude Fable 5.1Anthropic · Proprietary100 tasks / All-pass / five attempts26.2%
Claude Opus 5.5Anthropic · Proprietary100 tasks / All-pass / five attempts26.2%
Claude Sonnet 5.5Anthropic · Proprietary100 tasks / All-pass / five attempts25.8%
GPT-5.6 TerraOpenAI · Proprietary100 tasks / All-pass / five attempts24.6%
GPT-5.6 LunaOpenAI · Proprietary100 tasks / All-pass / five attempts24%
Grok 4.7xAI · Proprietary100 tasks / All-pass / five attempts23.2%
Claude Opus 4.8Anthropic · Proprietary100 tasks / All-pass / five attempts22.8%
Gemini 3.8 FlashGoogle · Proprietary100 tasks / All-pass / five attempts22.8%
GPT-6 LunaOpenAI · Proprietary100 tasks / All-pass / five attempts22.8%
GPT-5.5OpenAI · Proprietary100 tasks / All-pass / five attempts22.6%
Kimi K3Moonshot AI · Open source / open weights100 tasks / All-pass / five attempts22%
Gemini 4 ArgonGoogle · Proprietary100 tasks / All-pass / five attempts21.8%
Claude Opus 5Anthropic · Proprietary100 tasks / All-pass / five attempts21.6%
Claude Haiku 5.5Anthropic · Proprietary100 tasks / All-pass / five attempts20.8%
MiniMax M3.1 FlashMiniMax · Open source / open weights100 tasks / All-pass / five attempts20.8%
MiMo-V2.6-ProXiaomi · Open source / open weights100 tasks / All-pass / five attempts19.2%
Muse Spark OpenMeta · Open source / open weights100 tasks / All-pass / five attempts18.6%
Gemini 3.1 ProGoogle · Proprietary100 tasks / All-pass / five attempts17.8%
Muse Spark 1.2Meta · Proprietary100 tasks / All-pass / five attempts17.4%
DeepSeek V4.1 Flash 1017DeepSeek · Open source / open weights100 tasks / All-pass / five attempts17.2%
Qwen3.8 27BQwen · Open source / open weights100 tasks / All-pass / five attempts16.4%
Claude Sonnet 4.6Anthropic · Proprietary100 tasks / All-pass / five attempts15.8%
GLM-5.3-FlashZ.ai · Open source / open weights100 tasks / All-pass / five attempts15.4%
Step 5 PreviewStepFun · Open source / open weights100 tasks / All-pass / five attempts14.8%
Muse Spark 1.1Meta · Proprietary100 tasks / All-pass / five attempts14.4%
DeepSeek V4 Pro 0424DeepSeek · Open source / open weights100 tasks / All-pass / five attempts13.4%
GPT-5.4 miniOpenAI · Proprietary100 tasks / All-pass / five attempts13.4%
Claude Sonnet 5Anthropic · Proprietary100 tasks / All-pass / five attempts13.2%
Kimi K2.6Moonshot AI · Open source / open weights100 tasks / All-pass / five attempts13%
DeepSeek V4.1 FlashDeepSeek · Open source / open weights100 tasks / All-pass / five attempts12.8%
InklingThinking Machines · Open source / open weights100 tasks / All-pass / five attempts12.8%
Qwen3.5-397B-A17BQwen · Open source / open weights100 tasks / All-pass / five attempts12.2%
DeepSeek V4 Pro 0813DeepSeek · Open source / open weights100 tasks / All-pass / five attempts11.4%
GLM-5.3Z.ai · Open source / open weights100 tasks / All-pass / five attempts11.2%
DeepSeek V4 FlashDeepSeek · Open source / open weights100 tasks / All-pass / five attempts10.8%
DeepSeek V4 Flash 0420DeepSeek · Open source / open weights100 tasks / All-pass / five attempts10.8%
GLM-5.2Z.ai · Open source / open weights100 tasks / All-pass / five attempts10.4%
MiniMax M3MiniMax · Open source / open weights100 tasks / All-pass / five attempts9.8%
GLM-5.1Z.ai · Open source / open weights100 tasks / All-pass / five attempts8.4%
Nemotron 3.5 Super VLNVIDIA · Open source / open weights100 tasks / All-pass / five attempts8%
Qwen 3.5 9BQwen · Open source / open weights100 tasks / All-pass / five attempts6.6%
MiniMax M2.7MiniMax · Open source / open weights100 tasks / All-pass / five attempts6%
Claude Sonnet 4.5Anthropic · Proprietary100 tasks / All-pass / five attempts5.2%
gpt-oss-120bOpenAI · Open source / open weights100 tasks / All-pass / five attempts4%
DeepSeek R1 (Jan 2025)DeepSeek · Open source / open weights100 tasks / All-pass / five attempts2.4%
gpt-oss-20bOpenAI · Open source / open weights100 tasks / All-pass / five attempts2%
Qwen3-14BQwen · Open source / open weights100 tasks / All-pass / five attempts1.2%
View all 60 published reports
100 tasks / All-pass / five attempts

selection: best reported configuration per release · GPT-6 Astra (xhigh) · tools: benchmark-defined

100 tasks / All-pass / five attempts

selection: best reported configuration per release · GPT-6.1 Sol (high) · tools: benchmark-defined

100 tasks / All-pass / five attempts

selection: best reported configuration per release · Kimi K3.1 Pro · tools: benchmark-defined

100 tasks / All-pass / five attempts

selection: publisher default displayed configuration · GPT-6.1 Sol (max)

100 tasks / All-pass / five attempts

selection: publisher default displayed configuration · GPT-6 Astra (max)

100 tasks / All-pass / five attempts

selection: best reported configuration per release · Muse Wonder (max) · tools: benchmark-defined

100 tasks / All-pass / five attempts

selection: best reported configuration per release · Gemini 4 Carbon (medium) · tools: benchmark-defined

100 tasks / All-pass / five attempts

selection: best reported configuration per release · GPT-6.1 Terra (max) · tools: benchmark-defined

100 tasks / All-pass / five attempts

selection: publisher default displayed configuration · Gemini 4 Carbon (high)

100 tasks / All-pass / five attempts

selection: best reported configuration per release · GPT-5.6 Sol (high) · tools: benchmark-defined

100 tasks / All-pass / five attempts

selection: best reported configuration per release · Muse Spark 1.3 (max) · tools: benchmark-defined

100 tasks / All-pass / five attempts

selection: publisher default displayed configuration · Claude Fable 5.1 (max with fallback)

100 tasks / All-pass / five attempts

selection: publisher default displayed configuration · Claude Opus 5.5 (max with fallback)

100 tasks / All-pass / five attempts

selection: publisher default displayed configuration · Claude Sonnet 5.5 (max with fallback)

100 tasks / All-pass / five attempts

selection: best reported configuration per release · GPT-5.6 Terra (xhigh) · tools: benchmark-defined

100 tasks / All-pass / five attempts

selection: best reported configuration per release · GPT-5.6 Luna (max) · tools: benchmark-defined

100 tasks / All-pass / five attempts

selection: best reported configuration per release · Grok 4.7 (high) · tools: benchmark-defined

100 tasks / All-pass / five attempts

selection: best reported configuration per release · GPT-6 Luna (max) · tools: benchmark-defined

100 tasks / All-pass / five attempts

selection: best reported configuration per release · Claude Opus 4.8 (max) · tools: benchmark-defined

100 tasks / All-pass / five attempts

selection: best reported configuration per release · Gemini 3.8 Flash (medium) · tools: benchmark-defined

GPT-5.522.6%
100 tasks / All-pass / five attempts

selection: best reported configuration per release · GPT-5.5 (high) · tools: benchmark-defined

100 tasks / All-pass / five attempts

selection: best reported configuration per release · Kimi K3 (Max) · tools: benchmark-defined

100 tasks / All-pass / five attempts

selection: best reported configuration per release · Gemini 4 Argon (high) · tools: benchmark-defined

100 tasks / All-pass / five attempts

selection: best reported configuration per release · Claude Opus 5 (max) · tools: benchmark-defined

100 tasks / All-pass / five attempts

selection: publisher default displayed configuration · Gemini 3.8 Flash (high)

100 tasks / All-pass / five attempts

selection: best reported configuration per release · Claude Haiku 5.5 (max) · tools: benchmark-defined

100 tasks / All-pass / five attempts

selection: best reported configuration per release · MiniMax-M3.1-Flash · tools: benchmark-defined

100 tasks / All-pass / five attempts

selection: publisher default displayed configuration · Grok 4.7 (xhigh)

100 tasks / All-pass / five attempts

selection: best reported configuration per release · MiMo-V2.6-Pro · tools: benchmark-defined

100 tasks / All-pass / five attempts

selection: best reported configuration per release · Muse Spark Open (xhigh) · tools: benchmark-defined

100 tasks / All-pass / five attempts

selection: best reported configuration per release · Gemini 3.1 Pro Preview · tools: benchmark-defined

100 tasks / All-pass / five attempts

selection: best reported configuration per release · Muse Spark 1.2 (xhigh) · tools: benchmark-defined

100 tasks / All-pass / five attempts

selection: best reported configuration per release · DeepSeek V4.1 Flash 1017 (Max) · tools: benchmark-defined

100 tasks / All-pass / five attempts

selection: best reported configuration per release · Qwen3.8 27B (xhigh) · tools: benchmark-defined

100 tasks / All-pass / five attempts

selection: best reported configuration per release · Claude Sonnet 4.6 (max) · tools: benchmark-defined

100 tasks / All-pass / five attempts

selection: best reported configuration per release · GLM-5.3-Flash · tools: benchmark-defined

100 tasks / All-pass / five attempts

selection: best reported configuration per release · Step 5 Preview · tools: benchmark-defined

100 tasks / All-pass / five attempts

selection: best reported configuration per release · Muse Spark 1.1 (xhigh) · tools: benchmark-defined

100 tasks / All-pass / five attempts

selection: best reported configuration per release · GPT-5.4 mini (xhigh) · tools: benchmark-defined

100 tasks / All-pass / five attempts

selection: best reported configuration per release · DeepSeek V4 Pro 0424 (Max) · tools: benchmark-defined

100 tasks / All-pass / five attempts

selection: best reported configuration per release · Claude Sonnet 5 (max) · tools: benchmark-defined

100 tasks / All-pass / five attempts

selection: best reported configuration per release · Kimi K2.6 (Reasoning) · tools: benchmark-defined

Inkling12.8%
100 tasks / All-pass / five attempts

selection: best reported configuration per release · Inkling (xhigh) · tools: benchmark-defined

100 tasks / All-pass / five attempts

selection: best reported configuration per release · DeepSeek V4.1 Flash (Max) · tools: benchmark-defined

100 tasks / All-pass / five attempts

selection: best reported configuration per release · Qwen3.5 397B A17B · tools: benchmark-defined

100 tasks / All-pass / five attempts

selection: best reported configuration per release · DeepSeek V4 Pro 0813 (Max) · tools: benchmark-defined

GLM-5.311.2%
100 tasks / All-pass / five attempts

selection: best reported configuration per release · GLM-5.3 (Max) · tools: benchmark-defined

100 tasks / All-pass / five attempts

selection: best reported configuration per release · DeepSeek V4 Flash (max) · tools: benchmark-defined

100 tasks / All-pass / five attempts

selection: best reported configuration per release · DeepSeek V4 Flash 0420 (Max) · tools: benchmark-defined

GLM-5.210.4%
100 tasks / All-pass / five attempts

selection: best reported configuration per release · GLM-5.2 (Max) · tools: benchmark-defined

100 tasks / All-pass / five attempts

selection: best reported configuration per release · MiniMax-M3 · tools: benchmark-defined

100 tasks / All-pass / five attempts

selection: best reported configuration per release · GLM-5.1 · tools: benchmark-defined

100 tasks / All-pass / five attempts

selection: best reported configuration per release · Nemotron 3.5 Super VL · tools: benchmark-defined

100 tasks / All-pass / five attempts

selection: best reported configuration per release · Qwen3.5 9B · tools: benchmark-defined

100 tasks / All-pass / five attempts

selection: best reported configuration per release · MiniMax-M2.7 · tools: benchmark-defined

100 tasks / All-pass / five attempts

selection: best reported configuration per release · Claude 4.5 Sonnet · tools: benchmark-defined

100 tasks / All-pass / five attempts

selection: best reported configuration per release · gpt-oss-120b (High) · tools: benchmark-defined

100 tasks / All-pass / five attempts

selection: best reported configuration per release · DeepSeek R1 (Jan '25) · tools: benchmark-defined

100 tasks / All-pass / five attempts

selection: best reported configuration per release · gpt-oss-20b (High) · tools: benchmark-defined

100 tasks / All-pass / five attempts

selection: best reported configuration per release · Qwen3 14B · tools: benchmark-defined