LLM Reference
GPQAactiveReasoning

GPQA: Google-Proof Q&A

Metric: Accuracy (higher is better)Introduced: 2023

PhD-level multiple-choice questions in biology, physics, and chemistry designed so that even experts with internet access score below 67%. The Diamond subset (198 questions) is the hardest variant used in most frontier model evaluations. High benchmark score alone doesn't make a model the right pick — weigh it against pricing, API availability, and release date.

Models ranked

157

tracked on this benchmark

Score band

94.6 – 25.2

best → lowest tracked

Snapshot trend

+5.40

Jul 3 → Jul 16 · 1 models

Leaderboard

Tracked models ranked by Accuracy (higher is better).

Compare candidates
#Model variant and provenanceScore
1
Claude Mythos Preview
Version: diamondHarness: Not recordedEvaluator: Not recordedObserved: May 5, 2026Confidence: Not recordedSource 1Source 2

Notes: Confidence: medium. DAT-3780 weekly best-of refresh: multiple secondary trackers agree on 94.6%; pending an Anthropic primary score.

94.6
2
GPT-5.6 Sol
Version: GPQA Diamond; max reasoningHarness: Not recordedEvaluator: Not recordedObserved: Jul 9, 2026Confidence: Not recordedSource

Notes: Official GPT-5.6 GA launch benchmark row. Uses the repo canonical gpqa slug for the GPQA Diamond variant.

94.6
3
Gemini 3.1 Pro Preview
Version: diamondHarness: Not recordedEvaluator: Not recordedObserved: Apr 27, 2026Confidence: Not recordedSource
94.3
4
Claude Opus 4.7
Version: diamondHarness: Not recordedEvaluator: Not recordedObserved: Apr 28, 2026Confidence: Not recordedSource
94.2
5
GPT-5.5
Version: diamondHarness: Not recordedEvaluator: Not recordedObserved: Apr 23, 2026Confidence: Not recordedSource
93.6
6
GPT-5.5 Pro
Version: GPQA DiamondHarness: Not recordedEvaluator: Not recordedObserved: Apr 24, 2026Confidence: Not recordedSource

Notes: Uses the repo canonical gpqa slug for the GPQA Diamond variant; do not create a duplicate gpqa-diamond benchmark.

93.6
7
Claude Opus 4.8
Version: GPQA DiamondHarness: Not recordedEvaluator: Not recordedObserved: May 28, 2026Confidence: Not recordedSource

Notes: Official Anthropic launch benchmark using the repo canonical gpqa slug for GPQA Diamond.

93.6
8
Kimi K3
Version: GPQA-Diamond; max reasoning effortHarness: Not recordedEvaluator: Not recordedObserved: Jul 16, 2026Confidence: Not recordedSource

Notes: Model variant: Kimi K3 (max reasoning effort). Benchmark variant: GPQA-Diamond. Harness/evaluator not disclosed by Moonshot. Source posture: Moonshot/Kimi self-reported launch table; confidence: medium because the value is vendor-reported and not independently reproduced. Recommended seed value: 93.5.

93.5
9
MiniMax M3
Version: MiniMax M3 GPQA 92 (accuracy%)Harness: Not recordedEvaluator: Not recordedObserved: Jun 7, 2026Confidence: Not recordedSource
92.9
10
Qwen3.7-Max
Version: diamondHarness: Not recordedEvaluator: Not recordedObserved: May 20, 2026Confidence: Not recordedSource
92.4
11
Gemini 3.5 Flash
Version: GPQA Diamond (accuracy)Harness: Not recordedEvaluator: Not recordedObserved: Jun 7, 2026Confidence: Not recordedSource
92.2
12
GPT-5.4
Version: diamondHarness: Not recordedEvaluator: Not recordedObserved: Apr 18, 2026Confidence: Not recordedSource
92.0
13
Gemini 3 Pro
Version: Not recordedHarness: Not recordedEvaluator: Not recordedObserved: Dec 11, 2025Confidence: Not recordedSource
91.9
14
Qwen3.6-Max
Version: Not recordedHarness: Not recordedEvaluator: Not recordedObserved: Apr 13, 2026Confidence: Not recordedSource
91.8
15
Claude Opus 4.6
Version: diamondHarness: Not recordedEvaluator: Not recordedObserved: Apr 28, 2026Confidence: Not recordedSource
91.3
16
GLM-5.2
Version: diamondHarness: Not recordedEvaluator: Not recordedObserved: Jun 13, 2026Confidence: Not recordedSource
91.2
17
Kimi K2.6
Version: GPQA Diamond (accuracy)Harness: Not recordedEvaluator: Not recordedObserved: Jun 7, 2026Confidence: Not recordedSource
90.5
18
Gemini 3 Flash
Version: Not recordedHarness: Not recordedEvaluator: Not recordedObserved: Dec 17, 2025Confidence: Not recordedSource
90.4
19
Qwen3.6-Plus
Version: llm-stats shows 0 (accuracy%)Harness: Not recordedEvaluator: Not recordedObserved: Jun 7, 2026Confidence: Not recordedSource
90.4
20
DeepSeek V4 Pro
Version: diamondHarness: Not recordedEvaluator: Not recordedObserved: Apr 24, 2026Confidence: Not recordedSource
90.1
21
Grok 4.3
Version: Not recordedHarness: Not recordedEvaluator: Not recordedObserved: May 6, 2026Confidence: Not recordedSource
90.1
22
Claude Sonnet 4.6
Version: diamondHarness: Not recordedEvaluator: Not recordedObserved: Apr 28, 2026Confidence: Not recordedSource
89.9
23
Muse Spark
Version: diamondHarness: Not recordedEvaluator: Not recordedObserved: May 5, 2026Confidence: Not recordedSource 1Source 2

Notes: Confidence: medium. DAT-3780 weekly best-of refresh: secondary sources report about 89.5%; pending a Meta primary score.

89.5
24
Qwen3.5-397B-A17B
Version: diamondHarness: Not recordedEvaluator: Not recordedObserved: Apr 18, 2026Confidence: Not recordedSource: Artificial Analysis
89.3
25
Trinity-Large-Thinking
Version: diamondHarness: Not recordedEvaluator: Not recordedObserved: Apr 24, 2026Confidence: Not recordedSource
89.2
Showing the top 25 of 157 tracked models. Browse all models.

How to read this benchmark

This benchmark scores models where higher is better. Scores are useful for directional filtering and shortlisting — not for universal quality ranking. Prefer benchmarks closest to your workload, then validate the linked model pages for pricing, context window, and provider availability.

Trust this score when

  • There is a fresh timestamped snapshot (or multiple snapshots) for this benchmark.
  • The model list covers the same version family you can actually deploy today.
  • Top candidates overlap with your required routing and feature requirements.

Be cautious when

  • There is only one benchmark snapshot or the dataset appears stale.
  • The benchmark metric direction is opposite of your decision objective.
  • The score difference between options is narrow and likely within implementation variance.

FAQ

What does the GPQA benchmark measure?

PhD-level multiple-choice questions in biology, physics, and chemistry designed so that even experts with internet access score below 67%. The Diamond subset (198 questions) is the hardest variant used in most frontier model evaluations. On this page it lists 157 tracked model variants where higher is better.

Is a higher Google-Proof Q&A score always better?

For this benchmark, higher is better. A high score helps you shortlist, but confirm pricing, context window, and provider availability on each model page before committing — the top scorer is not always the right pick for your workload or budget.

How current is this Google-Proof Q&A data?

This benchmark was last reviewed on Apr 15, 2026. The tracked score average moved +5.40 points across the last 3 snapshots.

Related benchmarks

Last reviewed: Apr 15, 2026