Benchmark · Reasoning and knowledge

GPQA Diamond

Multiple-choice questions in biology, chemistry and physics written by PhD-level experts and hard for non-experts to answer even with the web; the Diamond subset is 198 questions. Epoch AI asks zero-shot with chain of thought and scores an answer only in its required format.

Score
% of questions correct, higher is better
Run by
Epoch AI
Status
Current, newest result 19 November 2026
Updated
as models are released
Ceiling
The best result is within five points of a perfect score, so the test no longer tells the top models apart
Data
the published file, CC BY 4.0, read 20 November 2026
Method
epoch.ai/benchmarks/gpqa-diamond

Results

45 models we list, each at its best setting, best first.

#ModelScoreRangeWho ran itSettingDate
1GPT-6.2 Astra AA96.8%Independenteffort max2026-10-27source
2Gemini 4 Hydrogen B96.8%Independenteffort xhigh2026-11-13source
3Gemini 4 Carbon B96.5%Independenteffort high2026-11-02source
4Gemini 4 Argon B96.4%Independenteffort high2026-10-13source
5Gemini 4 Neon B96.3%Independenteffort xhigh2026-11-19source
6GPT-6 Astra A95.8%Independenteffort max2026-08-30source
7Claude Sonnet 5.5 A95.6%Independenteffort max2026-09-29source
8GPT-6.1 Sol AA95.4%Independenteffort max2026-09-29source
8Gemini 3.8 Flash AA95.4%Independenteffort high2026-09-02source
10Gemini 3.1 Pro (preview) AA94.4%Independenteffort high2026-08-06source
11GPT-6.1 Terra AA94.3%Independenteffort max2026-10-26source
11GPT-6 Sol A94.3%Independenteffort max2026-09-22source
13Kimi K3.1 Pro D94.2%Independenteffort max2026-11-06source
14Claude Fable 5.2 A93.9%Independenteffort max2026-11-06source
15DeepSeek V4.1 Flash C93.2%Independenteffort max2026-10-19source
16Kimi K3 B93.1%Independenteffort max2026-07-16source
17Grok 4.7 BB92.7%Independenteffort xhigh2026-09-22source
18DeepSeek-V4-Flash-0731 C91.0%Independenteffort max2026-08-02source
19GLM-5.3 B90.9%Independenteffort max2026-08-24source
19MiniMax M3 C90.9%Independent2026-08-10source
19DeepSeek V4 Pro D90.9%Independenteffort high2026-08-06source
22Claude Opus 5.5 A90.6%Independenteffort max2026-09-22source
23GPT-6 Luna A90.5%Independenteffort max2026-09-22source
24Mistral Large 4 C90.2%Independenteffort high2026-11-02source
24GLM-5.3-Flash B90.2%Independenteffort max2026-08-26source
26Claude Haiku 5.5 A89.7%Independenteffort max2026-10-09source
27Grok 4.3 BB88.8%Independenteffort high2026-06-17source
28Muse Spark Open C88.6%Independenteffort xhigh2026-10-15source
29Inkling-Small B88.5%Independenteffort xhigh2026-08-14source
30Inkling B88.3%Independenteffort xhigh2026-08-05source
31Qwen3.5-397B-A17B B86.4%Independenteffort none2026-08-07source
32Qwen3.6-27B B85.9%Independent2026-08-07source
33Nemotron 3 Ultra 550B A55B B85.3%Independent2026-08-10source
34Qwen3.6-35B-A3B B84.8%Independenteffort none2026-08-07source
35Gemini 3.5 Flash-Lite AA83.3%Independenteffort high2026-08-06source
36Gemini 3.1 Flash-Lite B81.8%Independenteffort high2026-08-06source
37Qwen3.5-9B C79.0%Independent2026-08-27source
38Gemma 4 31B IT BB75.8%Independenteffort minimal2026-08-06source
38gpt-oss-120b BB75.8%Independenteffort high2025-12-11source
40Gemma 4 26B A4B IT BB73.2%Independenteffort minimal2026-08-06source
41gpt-oss-20b B60.8%Independenteffort medium2026-08-27source
42GLM-4.7-Flash B60.5%Independent2026-08-30source
43Qwen3-30B-A3B-Instruct-2507 C55.6%Independent2026-08-30source
44Llama 3.3 70B B47.4%Independent2025-01-27source
45Qwen3-4B-Instruct-2507 C45.8%Independent2026-08-28source
Every setting, 63 rows
ModelAs the leaderboard names itScoreWho ran itSettingDate
GPT-6.2 Astragpt-6.2-astra96.8%Independenteffort max2026-10-27source
Gemini 4 Hydrogengemini-4-hydrogen96.8%Independenteffort xhigh2026-11-13source
Gemini 4 Carbongemini-4-carbon96.5%Independenteffort high2026-11-02source
Gemini 4 Argongemini-4-argon96.4%Independenteffort high2026-10-13source
Gemini 4 Neongemini-4-neon96.3%Independenteffort xhigh2026-11-19source
GPT-6 Astragpt-6-astra95.8%Independenteffort max2026-08-30source
Claude Sonnet 5.5claude-sonnet-5-595.6%Independenteffort max2026-09-29source
GPT-6.1 Solgpt-6.1-sol95.4%Independenteffort max2026-09-29source
Gemini 3.8 Flashgemini-3.8-flash95.4%Independenteffort high2026-09-02source
Gemini 3.1 Pro (preview)gemini-3.1-pro-preview94.4%Independenteffort high2026-08-06source
GPT-6.1 Terragpt-6.1-terra94.3%Independenteffort max2026-10-26source
GPT-6 Solgpt-6-sol94.3%Independenteffort max2026-09-22source
Kimi K3.1 Prokimi-k3.1-pro-together94.2%Independenteffort max2026-11-06source
Gemini 3.1 Pro (preview)gemini-3.1-pro-preview94.1%Independent2026-02-20source
Claude Fable 5.2claude-fable-5-293.9%Independenteffort max2026-11-06source
DeepSeek V4.1 Flashdeepseek-v4.1-flash-101793.2%Independenteffort max2026-10-19source
Kimi K3kimi-k393.1%Independenteffort max2026-07-16source
Grok 4.7grok-4.792.7%Independenteffort xhigh2026-09-22source
Kimi K3kimi-k391.9%Independenteffort high2026-08-07source
DeepSeek-V4-Flash-0731deepseek-v4-flash-073191.0%Independenteffort max2026-08-02source
GLM-5.3glm-5.390.9%Independenteffort max2026-08-24source
MiniMax M3MiniMax-M390.9%Independent2026-08-10source
DeepSeek V4 Prodeepseek-v4-pro90.9%Independenteffort high2026-08-06source
Claude Opus 5.5claude-opus-5-590.6%Independenteffort max2026-09-22source
GPT-6 Lunagpt-6-luna90.5%Independenteffort max2026-09-22source
Mistral Large 4mistral-large-4-0-oss90.2%Independenteffort high2026-11-02source
GLM-5.3-Flashglm-5.3-flash90.2%Independenteffort max2026-08-26source
DeepSeek V4.1 Flashdeepseek-v4.1-flash89.8%Independenteffort max2026-10-09source
Claude Haiku 5.5claude-haiku-5-589.7%Independenteffort max2026-10-09source
DeepSeek V4 Prodeepseek-v4-pro89.7%Independenteffort max2026-06-16source
Grok 4.3grok-4.388.8%Independenteffort high2026-06-17source
Muse Spark OpenMuse-Spark-Open-816B-A42B88.6%Independenteffort xhigh2026-10-15source
Inkling-Smallinkling-small88.5%Independenteffort xhigh2026-08-14source
InklingInkling88.3%Independenteffort xhigh2026-08-05source
Qwen3.5-397B-A17Bqwen3.5-397b-a17b86.4%Independenteffort none2026-08-07source
Qwen3.5-397B-A17Bqwen3.5-397b-a17b85.9%Independent2026-08-07source
Qwen3.6-27Bqwen3.6-27b85.9%Independent2026-08-07source
Nemotron 3 Ultra 550B A55Bnemotron-3-ultra85.3%Independent2026-08-10source
Kimi K3kimi-k384.8%Independenteffort low2026-08-07source
Qwen3.6-27Bqwen3.6-27b84.8%Independenteffort none2026-08-07source
Qwen3.6-35B-A3Bqwen3.6-35b-a3b84.8%Independenteffort none2026-08-07source
Qwen3.6-35B-A3Bqwen3.6-35b-a3b83.8%Independent2026-08-07source
Gemini 3.5 Flash-Litegemini-3.5-flash-lite83.3%Independenteffort high2026-08-06source
Gemini 3.1 Flash-Litegemini-3.1-flash-lite81.8%Independenteffort high2026-08-06source
MiniMax M3MiniMax-M381.3%Independenteffort none2026-08-10source
Qwen3.5-9Bqwen3.5-9B79.0%Independent2026-08-27source
Qwen3.5-9Bqwen3.5-9B78.9%Independenteffort none2026-08-27source
Gemini 3.5 Flash-Litegemini-3.5-flash-lite75.8%Independenteffort low2026-08-06source
Gemma 4 31B ITgemma-4-31b-it75.8%Independenteffort minimal2026-08-06source
gpt-oss-120bopenai/gpt-oss-120b75.8%Independenteffort high2025-12-11source
Gemini 3.1 Flash-Litegemini-3.1-flash-lite74.2%Independenteffort low2026-08-06source
Gemini 3.5 Flash-Litegemini-3.5-flash-lite74.2%Independenteffort minimal2026-08-06source
Gemini 3.1 Flash-Litegemini-3.1-flash-lite73.7%Independenteffort minimal2026-08-06source
DeepSeek V4 Prodeepseek-v4-pro73.2%Independenteffort none2026-08-06source
Gemma 4 26B A4B ITgemma-4-26b-a4b-it73.2%Independenteffort minimal2026-08-06source
gpt-oss-20bgpt-oss-20b60.8%Independenteffort medium2026-08-27source
GLM-4.7-Flashglm-4.7-flash60.5%Independent2026-08-30source
Qwen3-30B-A3B-Instruct-2507qwen3-30b-a3b-instruct-250755.6%Independent2026-08-30source
gpt-oss-20bgpt-oss-20b53.2%Independenteffort low2026-08-27source
Llama 3.3 70BLlama-3.3-70B-Instruct47.4%Independent2025-01-27source
gpt-oss-20bgpt-oss-20b46.0%Independenteffort high2026-08-06source
Qwen3-4B-Instruct-2507qwen3-4b-instruct-250745.8%Independent2026-08-28source
GLM-4.7-Flashglm-4.7-flash45.1%Independenteffort none2026-08-30source

A rank is among the models we list with a result, not a place on the whole leaderboard, and a model offered by several hosts is shown once, on its own listing. These results don't move our grades. Data: Epoch AI, Capabilities and benchmarking (CC BY 4.0), CC BY 4.0.

For companies

Do agents find, use and choose your tools?

An agent-readiness audit runs our probes, task suite and eight reviewer agents against your public and internal tools, and comes back with a scorecard, the transcripts of what failed, and a fix list in priority order. From $2,500, re-run included. We never take payment to move a rank. We do help companies earn one.