GPQA Diamond (HELM)
As of 29 Oct 2026, the top score on GPQA Diamond (HELM) is 80.3% by Gemini 3 Pro (Google), ahead of GPT-5 at 79.1%, across 68 model configurations with a published result (HELM Capabilities, result dated 24 Nov 2025). GPQA Diamond is a set of 198 multiple-choice questions in biology, physics and chemistry written by PhD-level domain experts. The questions are designed to be "Google-proof": a skilled non-expert with unlimited web access still answers only about a third correctly, while experts in the field score around two-thirds. The Diamond subset keeps only the questions that experts answered correctly and non-experts got wrong. GPQA Diamond (HELM) is shown for reference only and does not count toward the BenchLeader Index.
Key facts (as of 29 Oct 2026)
- Gemini 3 Pro holds the top GPQA Diamond (HELM) score at 80.3% (HELM Capabilities, 24 Nov 2025).
- GPT-5 is second on GPQA Diamond (HELM) at 79.1% (HELM Capabilities, 24 Nov 2025).
- GPQA Diamond (HELM): Stanford CRFM's own GPQA Diamond run. Shown for reference; Epoch's run is in the index (HELM Capabilities).
- 68 model configurations have a published GPQA Diamond (HELM) result on BenchLeader (BenchLeader, 29 Oct 2026).
- GPQA Diamond (HELM) does not feed the BenchLeader Index; it is shown for reference (BenchLeader methodology).
Stanford CRFM's own GPQA Diamond run. Shown for reference; Epoch's run is in the index.
- Published by
- HELM Capabilities
- Category
- Reasoning
- Index weight
- Reference only
- Models
- 68
- Data as of
- 29 Oct 2026
- Newest result
- 24 Nov 2025
The publisher has posted no new results on this board since 24 Nov 2025. BenchLeader re-reads it every day, so newer models appear here as soon as they are published.
Stanford Center for Research on Foundation Models (crfm.stanford.edu), Apache-2.0 results.
What the test looks like
GPQA Diamond is a set of 198 multiple-choice questions in biology, physics and chemistry written by PhD-level domain experts. The questions are designed to be "Google-proof": a skilled non-expert with unlimited web access still answers only about a third correctly, while experts in the field score around two-thirds. The Diamond subset keeps only the questions that experts answered correctly and non-experts got wrong.
How it is scored
Accuracy with chain-of-thought prompting.
What to keep in mind
Stanford CRFM's own run of GPQA Diamond in HELM's fixed harness. Shown for reference; Epoch AI's run is counted in the index. See also the GPQA Diamond board.
- 1Gemini 3 Pro80.3%
- 2GPT-579.1%
- 3GPT-5 mini75.6%
- 4o375.3%
- 5Gemini 2.5 Pro74.9%
- 6o4-mini73.5%
- 7Grok 472.6%
- 8Qwen3 235B A22B 250772.6%
- 9Claude Opus 4 (thinking)70.9%
- 10Claude Sonnet 4 (thinking)70.6%
- 11Claude Sonnet 4.568.6%
- 12gpt-oss-120b68.4%
- 13GPT-5 nano67.9%
- 14Grok 3 mini67.5%
- 15Claude Opus 466.6%
| # | ||||
|---|---|---|---|---|
| 1 |  Gemini 3 Pro | 80.3% | 60.2 | 2025-11-24 |
| 2 |  GPT-5 | 79.1% | 57.7 | 2025-11-24 |
| 3 | GPT-5 mini | 75.6% | 56.4 | 2025-11-24 |
| 4 | o3 | 75.3% | 61.0 | 2025-11-24 |
| 5 | Gemini 2.5 Pro | 74.9% | 54.5 | 2025-11-24 |
| 6 | o4-mini | 73.5% | 56.9 | 2025-11-24 |
| 7 |  Grok 4 | 72.6% | 58.4 | 2025-11-24 |
| 8 |  Qwen3 235B A22B 2507open ↗ | 72.6% | 50.9 | 2025-11-24 |
| 9 |  Claude Opus 4thinking | 70.9% | 55.2 | 2025-11-24 |
| 10 | Claude Sonnet 4thinking | 70.6% | 53.3 | 2025-11-24 |
| 11 | Claude Sonnet 4.5 | 68.6% | 53.7 | 2025-11-24 |
| 12 | gpt-oss-120bopen ↗ | 68.4% | 45.5 | 2025-11-24 |
| 13 | GPT-5 nano | 67.9% | 55.9 | 2025-11-24 |
| 14 | Grok 3 mini | 67.5% | 52.3 | 2025-11-24 |
| 15 | Claude Opus 4 | 66.6% | 52.6 | 2025-11-24 |
| 16 |  DeepSeek R1 0528open ↗ | 66.6% | 49.4 | 2025-11-24 |
| 17 |  Palmyra X5 | 66.1% | – | 2025-11-24 |
| 18 | GPT-4.1 | 65.9% | 48.5 | 2025-11-24 |
| 19 |  Kimi K2open ↗ | 65.2% | 51.0 | 2025-11-24 |
| 20 | Grok 3 | 65.0% | 52.5 | 2025-11-24 |
| 21 |  Llama 4 Maverickopen ↗ | 65.0% | 42.2 | 2025-11-24 |
| 22 | Claude Sonnet 4 | 64.3% | 50.5 | 2025-11-24 |
| 23 | Qwen3-Next 80B A3Bthinkingopen ↗ | 63.0% | 52.3 | 2025-11-24 |
| 24 | Qwen3 235B A22Bopen ↗ | 62.3% | 50.0 | 2025-11-24 |
| 25 | GPT-4.1 mini | 61.4% | 44.8 | 2025-11-24 |
| 26 | Claude 3.7 Sonnet | 60.8% | 49.9 | 2025-11-24 |
| 27 | Claude Haiku 4.5 | 60.5% | 47.7 | 2025-11-24 |
| 28 |  GLM-4.5-Airopen ↗ | 59.4% | 46.8 | 2025-11-24 |
| 29 | gpt-oss-20bopen ↗ | 59.4% | 44.1 | 2025-11-24 |
| 30 | Claude 3.5 Sonnet | 56.5% | 47.1 | 2025-11-24 |
| 31 | Gemini 2.0 Flash | 55.6% | 43.5 | 2025-11-24 |
| 32 | DeepSeek V3open ↗ | 53.8% | 44.5 | 2025-11-24 |
| 33 | Gemini 1.5 Pro 002 | 53.4% | 41.5 | 2025-11-24 |
| 34 | Llama 3.1 405Bopen ↗ | 52.2% | 41.9 | 2025-11-24 |
| 35 | GPT-4o | 52.0% | 43.3 | 2025-11-24 |
| 36 |  Nova Premier | 51.8% | 46.1 | 2025-11-24 |
| 37 | Llama 4 Scoutopen ↗ | 50.7% | 38.6 | 2025-11-24 |
| 38 | GPT-4.1 nano | 50.7% | 37.0 | 2025-11-24 |
| 39 | Gemini 2.0 Flash-Lite | 50.0% | 48.4 | 2025-11-24 |
| 40 | Nova Pro | 44.6% | 40.9 | 2025-11-24 |
| 41 | GPT-5.1 | 44.2% | 56.8 | 2025-11-24 |
| 42 | Gemini 1.5 Flash 002 | 43.7% | 38.3 | 2025-11-24 |
| 43 |  Mistral Large 2.1open | 43.5% | 42.6 | 2025-11-24 |
| 44 | Qwen2.5 72Bopen ↗ | 42.6% | 43.2 | 2025-11-24 |
| 45 | Llama 3.1 70Bopen ↗ | 42.6% | 40.9 | 2025-11-24 |
| 46 | Palmyra Fin | 42.2% | – | 2025-11-24 |
| 47 | Nova Lite | 39.7% | 38.3 | 2025-11-24 |
| 48 | Palmyra X 004 | 39.5% | – | 2025-11-24 |
| 49 | Mistral Small 3.1open | 39.2% | 38.5 | 2025-11-24 |
| 50 | Gemini 2.5 Flash | 39.0% | 52.7 | 2025-11-24 |
| 51 | Nova Micro | 38.3% | 39.2 | 2025-11-24 |
| 52 |  Granite 4.0 Small | 38.3% | – | 2025-11-24 |
| 53 | GPT-4o mini | 36.8% | 35.4 | 2025-11-24 |
| 54 | Palmyra Med | 36.8% | – | 2025-11-24 |
| 55 | Claude 3.5 Haiku | 36.3% | 38.8 | 2025-11-24 |
| 56 | Qwen2.5 7B Instructopen ↗ | 34.1% | 39.9 | 2025-11-24 |
| 57 | Mixtral 8x22B Instructopen ↗ | 33.4% | 37.5 | 2025-11-24 |
| 58 | Granite 3.3 8Bopen ↗ | 32.5% | – | 2025-11-24 |
| 59 |  Olmo 2 13B November 2024 | 31.6% | – | 2025-11-24 |
| 60 | Gemini 2.5 Flash-Lite | 30.9% | 48.3 | 2025-11-24 |
More reasoning benchmarks · All LLM benchmarks
Related benchmarks
Top head-to-heads on GPQA Diamond (HELM)
Cite as: BenchLeader, “GPQA Diamond (HELM) leaderboard”, data as of 29 Oct 2026, https://www.benchleader.com/benchmarks/helm_gpqa.
GPQA Diamond (HELM): questions
- What does GPQA Diamond (HELM) measure?
- GPQA Diamond is a set of 198 multiple-choice questions in biology, physics and chemistry written by PhD-level domain experts. The questions are designed to be "Google-proof": a skilled non-expert with unlimited web access still answers only about a third correctly, while experts in the field score around two-thirds. The Diamond subset keeps only the questions that experts answered correctly and non-experts got wrong. Scores are reported in percent of tasks solved; higher is better.
- What is the top score on GPQA Diamond (HELM)?
- As of 29 Oct 2026, the highest score on GPQA Diamond (HELM) is 80.3% by Gemini 3 Pro (result dated 24 Nov 2025), ahead of GPT-5 at 79.1%. Scores come from HELM Capabilities and are re-read daily.
- Which AI model leads GPQA Diamond (HELM)?
- Gemini 3 Pro leads GPQA Diamond (HELM) with 80.3% as of 29 Oct 2026, ahead of GPT-5 at 79.1% (result dated 24 Nov 2025).
- How many models have GPQA Diamond (HELM) results?
- 68 model configurations have a GPQA Diamond (HELM) result on BenchLeader, all taken from HELM Capabilities.
- Who runs GPQA Diamond (HELM) and how often is it updated?
- GPQA Diamond (HELM) is published by HELM Capabilities. BenchLeader re-reads the published results once a day and records the date each result was published.
- Does GPQA Diamond (HELM) count toward the BenchLeader Index?
- No. GPQA Diamond (HELM) is shown for reference but left out of the composite index.