Best LLMs for scientific reasoning

GPQA Diamond

GPQA Diamond tests graduate-level scientific knowledge and reasoning with difficult multiple-choice questions in biology, chemistry, and physics. Diamond is the 198-question subset selected for high-quality expert agreement. The results here use the independently reported Diamond evaluation.

Updated: Nov 5, 2026Data as of: Nov 2, 2026

Widely usedSaturated

Official benchmark

GPQA Diamond 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 / creator	Version	Reported score
GPT-6 AstraOpenAI · Proprietary	Diamond	96.263%
Gemini 3.8 FlashGoogle · Proprietary	Diamond	95.253%
Gemini 3.1 ProGoogle · Proprietary	Diamond	94.141%
GPT-5.6 SolOpenAI · Proprietary	Diamond	94.141%
Muse Spark 1.3Meta · Proprietary	Diamond	94.141%
Claude Fable 5.1Anthropic · Proprietary	Diamond	93.737%
Claude Opus 5Anthropic · Proprietary	Diamond	93.737%
GPT-5.5OpenAI · Proprietary	Diamond	93.535%
Kimi K3Moonshot AI · Open source / open weights	Diamond	93.535%
MiniMax M3MiniMax · Open source / open weights	Diamond	92.929%
DeepSeek V4 Pro 0813DeepSeek · Open source / open weights	Diamond	92.828%
Claude Fable 5Anthropic · Proprietary	Diamond	92.626%
Claude Opus 4.8Anthropic · Proprietary	Diamond	92.02%
GLM-5.3Z.ai · Open source / open weights	Diamond	91.717%
Claude Opus 4.7Anthropic · Proprietary	Diamond	91.414%
GLM-5.3-FlashZ.ai · Open source / open weights	Diamond	91.212%
Claude Sonnet 5Anthropic · Proprietary	Diamond	91.111%
Kimi K2.6Moonshot AI · Open source / open weights	Diamond	91.111%
DeepSeek V4 Pro 0424DeepSeek · Open source / open weights	Diamond	90.505%
Muse Spark 1.2Meta · Proprietary	Diamond	90.404%
Muse Spark 1.1Meta · Proprietary	Diamond	89.798%
Claude Opus 4.6Anthropic · Proprietary	Diamond	89.596%
GLM-5.2Z.ai · Open source / open weights	Diamond	89.495%
DeepSeek V4 Flash 0420DeepSeek · Open source / open weights	Diamond	89.394%
Muse SparkMeta · Proprietary	Diamond	88.384%
Kimi K2.5Moonshot AI · Open source / open weights	Diamond	87.879%
Kimi K2 ThinkingMoonshot AI · Open source / open weights	Diamond	83.838%
gpt-oss-120bOpenAI · Open source / open weights	Diamond	78.182%
DeepSeek R1 (Jan 2025)DeepSeek · Open source / open weights	Diamond	70.808%
gpt-oss-20bOpenAI · Open source / open weights	Diamond	68.788%

View all 67 published reports

GPT-6 Astra96.263%

Diamond

GPT-6 Astra (Xhigh); 198-question Diamond subset; independent AA evaluation

GPT-6 Astra96.061%

Diamond

GPT-6 Astra (Max); 198-question Diamond subset; independent AA evaluation

Gemini 3.8 Flash95.253%

Diamond

Gemini 3.8 Flash (High); 198-question Diamond subset; independent AA evaluation

GPT-6 Astra94.949%

Diamond

GPT-6 Astra (High); 198-question Diamond subset; independent AA evaluation

GPT-5.6 Sol94.141%

Diamond

GPT-5.6 Sol (Max); 198-question Diamond subset; independent AA evaluation

Muse Spark 1.394.141%

Diamond

Muse Spark 1.3 (Xhigh); 198-question Diamond subset; independent AA evaluation

Gemini 3.1 Pro94.141%

Diamond

Gemini 3.1 Pro Preview; 198-question Diamond subset; independent AA evaluation

GPT-6 Astra93.939%

Diamond

GPT-6 Astra (Medium); 198-question Diamond subset; independent AA evaluation

Claude Opus 593.737%

Diamond

Claude Opus 5 (High); 198-question Diamond subset; independent AA evaluation

Claude Fable 5.193.737%

Diamond

Claude Fable 5.1 (Max, Default Fallback); 198-question Diamond subset; independent AA evaluation

Claude Opus 593.737%

Diamond

Claude Opus 5 (Xhigh); 198-question Diamond subset; independent AA evaluation

Kimi K393.535%

Diamond

Kimi K3 (Max); 198-question Diamond subset; independent AA evaluation

Muse Spark 1.393.535%

Diamond

Muse Spark 1.3 (Max); 198-question Diamond subset; independent AA evaluation

Gemini 3.8 Flash93.535%

Diamond

Gemini 3.8 Flash (Medium); 198-question Diamond subset; independent AA evaluation

GPT-5.593.535%

Diamond

GPT-5.5 (Xhigh); 198-question Diamond subset; independent AA evaluation

Claude Fable 5.193.434%

Diamond

Claude Fable 5.1 (Xhigh, Default Fallback); 198-question Diamond subset; independent AA evaluation

Claude Opus 593.232%

Diamond

Claude Opus 5 (Max); 198-question Diamond subset; independent AA evaluation

GPT-5.593.232%

Diamond

GPT-5.5 (High); 198-question Diamond subset; independent AA evaluation

GPT-5.6 Sol93.131%

Diamond

GPT-5.6 Sol (Xhigh); 198-question Diamond subset; independent AA evaluation

GPT-6 Astra93.131%

Diamond

GPT-6 Astra (Low); 198-question Diamond subset; independent AA evaluation

MiniMax M392.929%

Diamond

MiniMax-M3; 198-question Diamond subset; independent AA evaluation

GPT-5.6 Sol92.828%

Diamond

GPT-5.6 Sol (High); 198-question Diamond subset; independent AA evaluation

DeepSeek V4 Pro 081392.828%

Diamond

DeepSeek V4 Pro 0813 (Max); 198-question Diamond subset; independent AA evaluation

GPT-5.6 Sol92.626%

Diamond

GPT-5.6 Sol (Medium); 198-question Diamond subset; independent AA evaluation

GPT-5.592.626%

Diamond

GPT-5.5 (Medium); 198-question Diamond subset; independent AA evaluation

Claude Fable 592.626%

Diamond

Claude Fable 5 (Max, Opus 4.8 Fallback); 198-question Diamond subset; independent AA evaluation

Gemini 3.8 Flash92.02%

Diamond

Gemini 3.8 Flash (Low); 198-question Diamond subset; independent AA evaluation

Claude Opus 4.892.02%

Diamond

Claude Opus 4.8 (Max); 198-question Diamond subset; independent AA evaluation

Claude Opus 591.919%

Diamond

Claude Opus 5 (Medium); 198-question Diamond subset; independent AA evaluation

GLM-5.391.717%

Diamond

GLM-5.3 (Max); 198-question Diamond subset; independent AA evaluation

Claude Opus 4.791.414%

Diamond

Claude Opus 4.7 (Max); 198-question Diamond subset; independent AA evaluation

GLM-5.3-Flash91.212%

Diamond

GLM-5.3-Flash; 198-question Diamond subset; independent AA evaluation

Kimi K2.691.111%

Diamond

Kimi K2.6 (Reasoning); 198-question Diamond subset; independent AA evaluation

Claude Sonnet 591.111%

Diamond

Claude Sonnet 5 (Max); 198-question Diamond subset; independent AA evaluation

GPT-5.591.01%

Diamond

GPT-5.5 (Low); 198-question Diamond subset; independent AA evaluation

Claude Fable 5.190.606%

Diamond

Claude Fable 5.1 (High, Default Fallback); 198-question Diamond subset; independent AA evaluation

DeepSeek V4 Pro 042490.505%

Diamond

DeepSeek V4 Pro 0424 (High); 198-question Diamond subset; independent AA evaluation

Muse Spark 1.290.404%

Diamond

Muse Spark 1.2 (Xhigh); 198-question Diamond subset; independent AA evaluation

Muse Spark 1.189.798%

Diamond

Muse Spark 1.1 (Xhigh); 198-question Diamond subset; independent AA evaluation

GPT-5.6 Sol89.798%

Diamond

GPT-5.6 Sol (Low); 198-question Diamond subset; independent AA evaluation

Claude Opus 4.689.596%

Diamond

Claude Opus 4.6 (Max); 198-question Diamond subset; independent AA evaluation

GLM-5.289.495%

Diamond

GLM-5.2 (Max); 198-question Diamond subset; independent AA evaluation

DeepSeek V4 Flash 042089.394%

Diamond

DeepSeek V4 Flash 0420 (Max); 198-question Diamond subset; independent AA evaluation

Claude Opus 588.889%

Diamond

Claude Opus 5 (Low); 198-question Diamond subset; independent AA evaluation

DeepSeek V4 Pro 042488.788%

Diamond

DeepSeek V4 Pro 0424 (Max); 198-question Diamond subset; independent AA evaluation

Claude Fable 5.188.586%

Diamond

Claude Fable 5.1 (Medium, Default Fallback); 198-question Diamond subset; independent AA evaluation

Claude Opus 4.788.485%

Diamond

Claude Opus 4.7 (Non-reasoning, High); 198-question Diamond subset; independent AA evaluation

Muse Spark88.384%

Diamond

Muse Spark; 198-question Diamond subset; independent AA evaluation

Claude Fable 5.188.081%

Diamond

Claude Fable 5.1 (Low, Default Fallback); 198-question Diamond subset; independent AA evaluation

Kimi K2.587.879%

Diamond

Kimi K2.5 (Reasoning); 198-question Diamond subset; independent AA evaluation

DeepSeek V4 Flash 042086.667%

Diamond

DeepSeek V4 Flash 0420 (High); 198-question Diamond subset; independent AA evaluation

Kimi K384.242%

Diamond

Kimi K3 (Low); 198-question Diamond subset; independent AA evaluation

Claude Opus 4.684.04%

Diamond

Claude Opus 4.6 (Non-reasoning, High); 198-question Diamond subset; independent AA evaluation

Kimi K2 Thinking83.838%

Diamond

Kimi K2 Thinking; 198-question Diamond subset; independent AA evaluation

Claude Sonnet 580%

Diamond

Claude Sonnet 5 (Non-reasoning, High); 198-question Diamond subset; independent AA evaluation

GPT-5.6 Sol78.99%

Diamond

GPT-5.6 Sol (Non-reasoning); 198-question Diamond subset; independent AA evaluation

Kimi K2.578.889%

Diamond

Kimi K2.5 (Non-reasoning); 198-question Diamond subset; independent AA evaluation

Kimi K2.678.788%

Diamond

Kimi K2.6 (Non-reasoning); 198-question Diamond subset; independent AA evaluation

gpt-oss-120b78.182%

Diamond

gpt-oss-120b (High); 198-question Diamond subset; independent AA evaluation

GPT-5.576.768%

Diamond

GPT-5.5 (Non-reasoning); 198-question Diamond subset; independent AA evaluation

DeepSeek V4 Pro 042471.717%

Diamond

DeepSeek V4 Pro 0424 (Non-reasoning); 198-question Diamond subset; independent AA evaluation

DeepSeek V4 Flash 042071.616%

Diamond

DeepSeek V4 Flash 0420 (Non-reasoning); 198-question Diamond subset; independent AA evaluation

DeepSeek R1 (Jan 2025)70.808%

Diamond

DeepSeek R1 (Jan '25); 198-question Diamond subset; independent AA evaluation

gpt-oss-20b68.788%

Diamond

gpt-oss-20b (High); 198-question Diamond subset; independent AA evaluation

GLM-5.268.586%

Diamond

GLM-5.2 (Non-reasoning); 198-question Diamond subset; independent AA evaluation

gpt-oss-120b67.172%

Diamond

gpt-oss-120b (Low); 198-question Diamond subset; independent AA evaluation

gpt-oss-20b61.111%

Diamond

gpt-oss-20b (Low); 198-question Diamond subset; independent AA evaluation

Other benchmarks

LiveBench: Reasoning

LiveBench 2.0 evaluates reasoning with LiveBench on its October 25, 2026 release.

LifeVal

LifeVal evaluates personal assistants on everyday responsibilities, including caregiving, household administration, shopping, education, personal finances, and accessing government services.

Gaia2

Gaia2 evaluates assistants coordinating tasks across simulated email, calendars, shopping, contacts, and other apps.