Where GPQA remains difficult

Aggregate question difficulty across 243 accepted configurations,
without publishing prompts, answer keys, or a model-by-question matrix.

Results 0.1.0Data through Aug 17, 2026

189 retained questions · 3 domains · 9 reportable subcategories · Aggregate-first publication boundary
Method notes

These controls preview a future, separately reviewed question-analysis release. This release intentionally publishes one aggregate cohort only.

Displayed statistics require ≥5 questions and ≥5 configurations.

1. Difficulty by domain

  1. Biology1770.0%70.0%
  2. Chemistry8860.9%60.9%
  3. Physics8478.9%78.9%

Each configuration is weighted equally. Repeated attempts reduce to one per-question mean before aggregation.

2. Subcategories with enough questions to report

4 subcategories suppressed: fewer than 5 questions
SubcategoryQuestionsMean correctnessValue
Astrophysics1371.7%71.7%
Chemistry (general)1969.8%69.8%
Electromagnetism and Photonics671.9%71.9%
High-energy particle physics1482.7%82.7%
Molecular Biology1470.5%70.5%
Organic Chemistry6858.3%58.3%
Physics (general)1878.1%78.1%
Quantum Mechanics2581.8%81.8%
Relativistic Mechanics683.2%83.2%

3. Consistently difficult patterns

1

Organic Chemistry is the hardest reportable subcategory.

58.3% mean correctness.

2

Difficulty remains broad.

The release does not expose Record IDs or per-model correctness.

3

Descriptive, not causal.

These aggregates do not identify why a configuration failed a question.

4. Question disclosure boundary

Included in this release

Domain/subcategory aggregates, counts, definitions, cohort scope, and suppression markers.

Held back pending dedicated review

Record IDs, per-model correctness, prompt text, choices, keys, completions, and reasoning.

5. How to read this page

  1. Attempts per configuration
  2. Per-question configuration mean
  3. Aggregate configurations equally
  4. Suppress small cells before rendering
Download aggregate JSON

Explore the benchmark

Open leaderboard See cleaning impact Read methodology