UPSC Bench
The UPSC Civil Services Examination is India's most competitive public exam and one of the toughest selection processes in the world. Each year, over 10 lakh (1 million) candidates register for the Preliminary round alone — a pair of multiple-choice papers in General Studies and aptitude — competing for roughly 1,000 positions. Most aspirants spend two to three years in full-time preparation, and the selection rate is under 0.1%.
Those who clear the exam join the Indian Administrative Service (IAS), Indian Police Service (IPS), Indian Foreign Service (IFS), and other elite branches of government. These officers wield extraordinary authority: a single IAS officer may govern a district of several million people as District Magistrate, shape national policy from the Central Secretariat, or oversee billions in public spending. The civil services form the backbone of Indian governance, and this exam has been the sole gateway into them since independence.
UPSC Bench evaluates frontier AI models against both stages of the exam: the objective Prelims (MCQ papers with negative marking) and the subjective Mains (essay and long-form answer papers graded by rubric). We estimate where each model would rank among the real candidate pool using historical score distributions.
Rankings: 2025
| # | Model | Score |
|---|---|---|
| 1 | GPT-5.2 | 897.3/125072% ✓ Pass |
| 2 | Gemini 3.1 Pro | 865.1/125069% ✓ Pass |
| 3 | Claude Opus 4.6 | 828.4/125066% ✓ Pass |
| 4 | Gemini 3 Flash | 780.9/125062% ✓ Pass |
| 5 | Gemini 2.5 Flash | 735.8/125059% ✓ Pass |
Shakti DubeyCSE 2024 AIR 1 · Est. 602/1250 (from 843/1750) | 602.1/125048% ✓ Pass |
Paper breakdown: Essay (250) · GS-I (250) · GS-II (250) · GS-III (250) · GS-IV (250). Essay scored as best 1 from each section (A & B).
Est. AIR = Estimated All-India Rank among ~14.5K Mains candidates. Scores graded by LLM-as-judge (rubric-based).
To pass (2025): Mains written score must exceed proportional cutoff of ~571/1250 (scaled from 800/1750 full exam cutoff).
Human reference: Shakti Dubey (CSE 2024 AIR 1, written 843/1750). Score proportionally estimated for our 1250-mark subset — UPSC does not publish paper-wise marks. Optional paper excluded from benchmark.
How we score
Each model writes full answers to all 87 Mains questions (8 essays + 79 GS). Answers are graded by a calibrated LLM judge (Claude Opus) using a 5-dimension rubric with UPSC-realistic score anchors. All 4 candidates for each question are graded comparatively to ensure differentiation. Essay scoring picks the best answer from each section (A & B).
Test yourself
Try 5 real GS Paper I questions from the 2025 exam. Get instant feedback, see your extrapolated score, and find out where you'd rank among AI models.
Marking Scheme
Prelims
GS Paper I
CSAT Paper II
Mains
Prelims: GS-I must exceed year-specific cutoff. CSAT is qualifying only (33% minimum). Mains: Answers graded by LLM judge on 5-dimension rubric. Cutoff scaled proportionally from full exam (800/1750 → 571/1250).