GPQA Diamond Benchmark
The gold standard for measuring AI scientific reasoning capability
GPQA Diamond (Graduate-level Google-Proof Q&A Diamond) is a benchmark that tests AI systems on expert-level questions in biology, chemistry, and physics. The "Diamond" subset contains the most challenging questions requiring deep scientific reasoning that cannot be answered through web search.
GPQA Diamond Leaderboard (July 2026)
| Rank | Model | Accuracy | Type |
|---|---|---|---|
| 1 | Gemini 3.1 Pro | 94.1% | General-purpose LLM |
| 2 | GPT-5.5 | 93.5% | General-purpose LLM |
| 3 | Omic AI Scientist | 93.3% | Specialized for biology & chemistry |
| 4 | Claude Fable 5 | 92.6% | General-purpose LLM |
| N/A | Human PhD Expert | 69.7% | Domain expert baseline |
Why GPQA Diamond Matters for Drug Discovery
Drug discovery requires expert-level reasoning in biology, chemistry, and pharmacology. GPQA Diamond measures exactly this capability. AI systems that score highly on GPQA Diamond can:
Understand Disease Biology
Reason about molecular mechanisms, pathway dysregulation, and disease etiology at PhD level.
Predict Molecular Interactions
Understand protein-ligand binding, enzyme kinetics, and chemical reactivity.
Analyze Experimental Data
Interpret complex biological datasets and draw valid scientific conclusions.
Design Therapeutic Strategies
Propose rational drug design approaches based on mechanistic understanding.
About the Benchmark
Question Characteristics
- • Graduate-level difficulty (PhD qualifying exam level)
- • Cannot be answered through web search
- • Require multi-step reasoning
- • Cover biology, chemistry, and physics
- • Vetted by domain experts
Evaluation Methodology
- • Multiple choice format for objective scoring
- • "Diamond" subset is most challenging tier
- • Human expert baseline established
- • Regularly updated with new questions
- • Prevents memorization-based performance
Why Omic Excels on GPQA Diamond
Omic's AI Scientist achieves a best-in-class 93.3% on GPQA Diamond, on par with frontier general models and far ahead of human PhD experts (69.7%). This performance comes from:
Focused on biology, chemistry, and drug discovery rather than general knowledge.
Deep understanding of genomics, proteomics, and metabolomics relationships.
Trained on disease mechanisms, not just isolated facts.
Frequently Asked Questions
What is a good GPQA Diamond score?
Human PhD experts score around 69.7%. Scores above 70% indicate PhD-level performance; above 90% indicates reasoning far ahead of human PhD experts. Top AI systems now exceed 90%.
Who has the highest GPQA Diamond score?
As of July 2026, frontier general models lead: Gemini 3.1 Pro (94.1%) and GPT-5.5 (93.5%). Omic AI Scientist follows closely with a best-in-class 93.3%, ahead of Claude Fable 5 (92.6%) and far ahead of human PhD experts (69.7%).
How is GPQA different from other benchmarks?
Unlike general knowledge benchmarks, GPQA specifically tests graduate-level scientific reasoning. Questions cannot be answered through memorization or web search, making it relevant for evaluating AI for scientific research.