SAS-Bench
Models grade short student answers from subject exams. The benchmark compares their scores and error classifications with expert annotations.
- Made by
- Peichao Lai and colleagues
- Released
- May 2025
- Kind
- Other professions
- Level
- University
- Scored
- Checked automatically
- Answers
- Short answer
- Tasks
- All public
- Measures
- Capability
- Subject
- Education
- Interaction
- Single turn
- Items from
- Exams and contests, Real work and use
- Refresh
- One fixed set
- Run by
- Anyone
- Strongest model tested
-
Qwen3-32B #135 on the capability index
- Last model added
- April 2025
- Scores refreshed
- November 20th, 2026
Deepseek-V3 (snapshot unspecified) tops the leaderboard with 74.1% Qwen3-32B leads with 67.2% among models released by April 2025
Best score per entry from the benchmark's paper, read on November 20th, 2026. A paper’s table doesn’t add newer models. Where a model ran with several settings or harnesses, we keep its best. Names in gray aren't in our model list: fine-tunes, agent systems, products and older models. They don't count toward leaders. The best model we track is Qwen3-32B, at 67.2%. This view shows only models we track, each at its best.
The newest model with a score here came out in April 2025, and nobody has tested a newer one. Among the 20 highest-ranked models on the capability index, these have no score yet:
- GPT-6.2 Astra
- Claude Fable 5.2
-
Gemini 4 Carbon
- GPT-6 Astra
- and 16 more
- 74.1 Deepseek-V3 (snapshot unspecified)
- 73.8 Deepseek-R1 (snapshot unspecified)
- 67.2
Qwen3-32B
- 65.2 TinyR1-32B-Preview
- 64.5
QwQ-32B
- 64.4 Qwen2.5-14B-Instruct
- 62.7
Llama 3.3 70B Instruct
- 62.6 Qwen2.5-32B-Instruct
- 58.6 GPT-4o mini
- 58.4
Qwen3-8B
- 46.9 GLM-4-9B-Chat
- 46.0 MiMo-7B-RL
- 40.4
DeepSeek-R1-Distill-Qwen-7B
- 36.1 Llama3-8B-Instruct
- 35.5 Mixtral 8×7B-Instruct
- 21.6 Deepseek-Prover-V2-7B
- 1 Deepseek-V3 (snapshot unspecified) 74.1%
- 2 Deepseek-R1 (snapshot unspecified) 73.8%
- 3
Qwen3-32B 67.2%
- 4 TinyR1-32B-Preview 65.2%
- 5
QwQ-32B 64.5%
- 6 Qwen2.5-14B-Instruct 64.4%
- 7
Llama 3.3 70B Instruct 62.7%
- 8 Qwen2.5-32B-Instruct 62.6%
- 9 GPT-4o mini 58.6%
- 10
Qwen3-8B 58.4%
- 11 GLM-4-9B-Chat 46.9%
- 12 MiMo-7B-RL 46%
- 13
DeepSeek-R1-Distill-Qwen-7B 40.4%
- 14 Llama3-8B-Instruct 36.1%
- 15 Mixtral 8×7B-Instruct 35.5%
- 16 Deepseek-Prover-V2-7B 21.6%
Source: The SAS-Bench paper
- 1
Qwen3-32B 67.2%
- 2
QwQ-32B 64.5%
- 3
Llama 3.3 70B Instruct 62.7%
- 4 GPT-4o mini 58.6%
- 5
Qwen3-8B 58.4%
- 6 GLM-4-9B-Chat 46.9%
- 7 MiMo-7B-RL 46%
- 8
DeepSeek-R1-Distill-Qwen-7B 40.4%
Source: The SAS-Bench paper
| # | Entry | Score | Setting | Date | Source |
|---|---|---|---|---|---|
| 1 | Deepseek-V3 (snapshot unspecified) | 74.1% | Official leaderboard | ||
| 2 | Deepseek-R1 (snapshot unspecified) | 73.8% | Official leaderboard | ||
| 3 | | 67.2% | Apr 2025 | Official leaderboard | |
| 4 | TinyR1-32B-Preview | 65.2% | Official leaderboard | ||
| 5 | | 64.5% | Mar 2025 | Official leaderboard | |
| 6 | Qwen2.5-14B-Instruct | 64.4% | Official leaderboard | ||
| 7 | | 62.7% | Dec 2024 | Official leaderboard | |
| 8 | Qwen2.5-32B-Instruct | 62.6% | Official leaderboard | ||
| 9 | GPT-4o mini | 58.6% | Jul 2024 | Official leaderboard | |
| 10 | | 58.4% | Apr 2025 | Official leaderboard | |
| 11 | GLM-4-9B-Chat | 46.9% | Jun 2024 | Official leaderboard | |
| 12 | MiMo-7B-RL | 46% | Apr 2025 | Official leaderboard | |
| 13 | | 40.4% | Jan 2025 | Official leaderboard | |
| 14 | Llama3-8B-Instruct | 36.1% | Official leaderboard | ||
| 15 | Mixtral 8×7B-Instruct | 35.5% | Official leaderboard | ||
| 16 | Deepseek-Prover-V2-7B | 21.6% | Official leaderboard |
| # | Model | Score | Setting | Date | Source |
|---|---|---|---|---|---|
| 1 | | 67.2% | Apr 2025 | Official leaderboard | |
| 2 | | 64.5% | Mar 2025 | Official leaderboard | |
| 3 | | 62.7% | Dec 2024 | Official leaderboard | |
| 4 | GPT-4o mini | 58.6% | Jul 2024 | Official leaderboard | |
| 5 | | 58.4% | Apr 2025 | Official leaderboard | |
| 6 | GLM-4-9B-Chat | 46.9% | Jun 2024 | Official leaderboard | |
| 7 | MiMo-7B-RL | 46% | Apr 2025 | Official leaderboard | |
| 8 | | 40.4% | Jan 2025 | Official leaderboard |
Rankings by part
We read 24 parts from the benchmark’s paper in 2 groups: ccs and ecs. Each ranks models on one part of the test. QwQ-32B leads 8 of them, and 6 other models lead the rest. Arrows show how far a model moved from its overall rank among the models we track.
Phy. (S.)
8 models we track have a score here.
- 1 GPT-4o mini ↑3 58.9%
- 2
Qwen3-8B ↑3 54.3%
- 3 MiMo-7B-RL ↑4 52.8%
- 4
QwQ-32B ↓2 48.5%
- 5
Qwen3-32B ↓4 47.3%
- 6 GLM-4-9B-Chat 45.6%
- 7
Llama 3.3 70B Instruct ↓4 45.3%
- 8
DeepSeek-R1-Distill-Qwen-7B 33.7%
Phy. (M.)
8 models we track have a score here.
- 1
QwQ-32B ↑1 87.2%
- 2
Qwen3-32B ↓1 85.5%
- 3 GPT-4o mini ↑1 81.2%
- 4
Qwen3-8B ↑1 76.2%
- 5
Llama 3.3 70B Instruct ↓2 70%
- 6 GLM-4-9B-Chat 52.3%
- 7 MiMo-7B-RL 41%
- 8
DeepSeek-R1-Distill-Qwen-7B 29.2%
His. (S.)
8 models we track have a score here.
- 1
QwQ-32B ↑1 75.4%
- 2
Llama 3.3 70B Instruct ↑1 72%
- 3
Qwen3-32B ↓2 65%
- 4 MiMo-7B-RL ↑3 61.3%
- 5 GPT-4o mini ↓1 54.9%
- 6
DeepSeek-R1-Distill-Qwen-7B ↑2 50.9%
- 7
Qwen3-8B ↓2 45.5%
- 8 GLM-4-9B-Chat ↓2 36.8%
Geo. (S.)
8 models we track have a score here.
- 1
Qwen3-32B 80.4%
- 2
QwQ-32B 77.1%
- 3 GPT-4o mini ↑1 76.6%
- 4
Llama 3.3 70B Instruct ↓1 72.5%
- 5 GLM-4-9B-Chat ↑1 69.4%
- 6
Qwen3-8B ↓1 68.9%
- 7 MiMo-7B-RL 67.1%
- 8
DeepSeek-R1-Distill-Qwen-7B 32.4%
Bio. (G.)
8 models we track have a score here.
- 1
QwQ-32B ↑1 72.5%
- 2
Llama 3.3 70B Instruct ↑1 67.9%
- 3 GPT-4o mini ↑1 65.4%
- 4
Qwen3-32B ↓3 63.1%
- 5
DeepSeek-R1-Distill-Qwen-7B ↑3 52.2%
- 6
Qwen3-8B ↓1 43.2%
- 7 GLM-4-9B-Chat ↓1 39.2%
- 8 MiMo-7B-RL ↓1 35.9%
Chi. (G.)
8 models we track have a score here.
- 1
QwQ-32B ↑1 96%
- 2
Qwen3-32B ↓1 92.2%
- 3 GPT-4o mini ↑1 87.7%
- 4
Qwen3-8B ↑1 86%
- 5
Llama 3.3 70B Instruct ↓2 85.3%
- 6 GLM-4-9B-Chat 63.9%
- 7 MiMo-7B-RL 54.7%
- 8
DeepSeek-R1-Distill-Qwen-7B 52.4%
Chi. (S.)
8 models we track have a score here.
- 1 GPT-4o mini ↑3 55.3%
- 2
Qwen3-32B ↓1 50.4%
- 3
DeepSeek-R1-Distill-Qwen-7B ↑5 44.3%
- 4 MiMo-7B-RL ↑3 43.1%
- 5 GLM-4-9B-Chat ↑1 42.9%
- 6
Qwen3-8B ↓1 42%
- 7
Llama 3.3 70B Instruct ↓4 35.8%
- 8
QwQ-32B ↓6 31.8%
Math (S.)
8 models we track have a score here.
- 1
Llama 3.3 70B Instruct ↑2 58.6%
- 2
Qwen3-32B ↓1 51.3%
- 3
QwQ-32B ↓1 48.7%
- 4
Qwen3-8B ↑1 46.3%
- 5 GPT-4o mini ↓1 43.6%
- 6 MiMo-7B-RL ↑1 38.1%
- 7 GLM-4-9B-Chat ↓1 35.5%
- 8
DeepSeek-R1-Distill-Qwen-7B 29.5%
Math (G.)
8 models we track have a score here.
- 1
Qwen3-32B 80.8%
- 2
Llama 3.3 70B Instruct ↑1 75%
- 3
Qwen3-8B ↑2 73.3%
- 4 GLM-4-9B-Chat ↑2 57%
- 5 MiMo-7B-RL ↑2 55.8%
- 6
QwQ-32B ↓4 45.5%
- 7
DeepSeek-R1-Distill-Qwen-7B ↑1 39.5%
- 8 GPT-4o mini ↓4 37.4%
Pol. (S.)
8 models we track have a score here.
- 1
QwQ-32B ↑1 74.5%
- 2
Qwen3-32B ↓1 73.3%
- 3
Qwen3-8B ↑2 64.3%
- 4
Llama 3.3 70B Instruct ↓1 63.7%
- 5 GPT-4o mini ↓1 63.4%
- 6 GLM-4-9B-Chat 54.8%
- 7
DeepSeek-R1-Distill-Qwen-7B ↑1 53.8%
- 8 MiMo-7B-RL ↓1 36.8%
Eng. (G.)
8 models we track have a score here.
- 1
Llama 3.3 70B Instruct ↑2 67.6%
- 2
Qwen3-32B ↓1 59.3%
- 3
QwQ-32B ↓1 54.8%
- 4
Qwen3-8B ↑1 50.5%
- 5 MiMo-7B-RL ↑2 34.7%
- 6 GLM-4-9B-Chat 33.9%
- 7
DeepSeek-R1-Distill-Qwen-7B ↑1 33%
- 8 GPT-4o mini ↓4 22.6%
Che. (G.)
8 models we track have a score here.
- 1
QwQ-32B ↑1 62.2%
- 2
Qwen3-32B ↓1 57.8%
- 3 GPT-4o mini ↑1 56%
- 4
Qwen3-8B ↑1 50.5%
- 5
Llama 3.3 70B Instruct ↓2 38.9%
- 6
DeepSeek-R1-Distill-Qwen-7B ↑2 34.3%
- 7 MiMo-7B-RL 31.1%
- 8 GLM-4-9B-Chat ↓2 30.8%
Phy. (S.)
8 models we track have a score here.
- 1 MiMo-7B-RL ↑6 51%
- 2 GLM-4-9B-Chat ↑4 35%
- 3
Qwen3-8B ↑2 23.4%
- 4
Llama 3.3 70B Instruct ↓1 22.6%
- 5 GPT-4o mini ↓1 17.9%
- 6
QwQ-32B ↓4 4.74%
- 7
Qwen3-32B ↓6 -4.17%
- 8
DeepSeek-R1-Distill-Qwen-7B -45.19%
Phy. (M.)
8 models we track have a score here.
- 1
QwQ-32B ↑1 63.9%
- 2
Qwen3-8B ↑3 63.9%
- 3
Llama 3.3 70B Instruct 57.4%
- 4 GLM-4-9B-Chat ↑2 24.2%
- 5 GPT-4o mini ↓1 24.2%
- 6
DeepSeek-R1-Distill-Qwen-7B ↑2 24.2%
- 7
Qwen3-32B ↓6 24.2%
- 8 MiMo-7B-RL ↓1 24.2%
His. (S.)
8 models we track have a score here.
- 1
Qwen3-32B 69.5%
- 2
QwQ-32B 67.1%
- 3 GPT-4o mini ↑1 62.1%
- 4
Llama 3.3 70B Instruct ↓1 54.3%
- 5 GLM-4-9B-Chat ↑1 32.5%
- 6
Qwen3-8B ↓1 14.3%
- 7 MiMo-7B-RL 14.3%
- 8
DeepSeek-R1-Distill-Qwen-7B 0.95%
Geo. (S.)
8 models we track have a score here.
- 1
QwQ-32B ↑1 70%
- 2
Qwen3-32B ↓1 54.3%
- 3
Llama 3.3 70B Instruct 42.1%
- 4 MiMo-7B-RL ↑3 38.9%
- 5 GPT-4o mini ↓1 36.7%
- 6 GLM-4-9B-Chat 34.7%
- 7
Qwen3-8B ↓2 -4.96%
- 8
DeepSeek-R1-Distill-Qwen-7B -38.66%
Bio. (G.)
8 models we track have a score here.
- 1 GLM-4-9B-Chat ↑5 62.1%
- 2 MiMo-7B-RL ↑5 58.4%
- 3 GPT-4o mini ↑1 55.2%
- 4
QwQ-32B ↓2 53.7%
- 5
Qwen3-32B ↓4 53.7%
- 6
Qwen3-8B ↓1 52.2%
- 7
Llama 3.3 70B Instruct ↓4 45.1%
- 8
DeepSeek-R1-Distill-Qwen-7B 23.6%
Chi. (G.)
8 models we track have a score here.
- 1 MiMo-7B-RL ↑6 92.2%
- 2 GPT-4o mini ↑2 79%
- 3
Llama 3.3 70B Instruct 52.7%
- 4
Qwen3-32B ↓3 52.7%
- 5
QwQ-32B ↓3 51.1%
- 6
Qwen3-8B ↓1 47.8%
- 7 GLM-4-9B-Chat ↓1 20.4%
- 8
DeepSeek-R1-Distill-Qwen-7B -20.36%
Chi. (S.)
8 models we track have a score here.
- 1 GPT-4o mini ↑3 78%
- 2 GLM-4-9B-Chat ↑4 77.3%
- 3
QwQ-32B ↓1 69.2%
- 4 MiMo-7B-RL ↑3 63.1%
- 5
Qwen3-32B ↓4 47.3%
- 6
Llama 3.3 70B Instruct ↓3 46.3%
- 7
Qwen3-8B ↓2 34%
- 8
DeepSeek-R1-Distill-Qwen-7B 3.87%
Math (S.)
8 models we track have a score here.
- 1
Qwen3-32B 82.2%
- 2
QwQ-32B 79%
- 3 GPT-4o mini ↑1 67.6%
- 4 GLM-4-9B-Chat ↑2 63.8%
- 5
Llama 3.3 70B Instruct ↓2 54.3%
- 6
Qwen3-8B ↓1 39.2%
- 7 MiMo-7B-RL 13.4%
- 8
DeepSeek-R1-Distill-Qwen-7B -23.81%
Math (G.)
8 models we track have a score here.
- 1 GLM-4-9B-Chat ↑5 46.9%
- 2 GPT-4o mini ↑2 46.9%
- 3 MiMo-7B-RL ↑4 35.1%
- 4
Llama 3.3 70B Instruct ↓1 30%
- 5
Qwen3-32B ↓4 18.3%
- 6
QwQ-32B ↓4 16.8%
- 7
Qwen3-8B ↓2 -8.14%
- 8
DeepSeek-R1-Distill-Qwen-7B -13.57%
Pol. (S.)
8 models we track have a score here.
- 1 GPT-4o mini ↑3 92.3%
- 2 GLM-4-9B-Chat ↑4 82.4%
- 3
Qwen3-32B ↓2 62.1%
- 4
Llama 3.3 70B Instruct ↓1 58.8%
- 5
Qwen3-8B 57.2%
- 6
QwQ-32B ↓4 48.8%
- 7
DeepSeek-R1-Distill-Qwen-7B ↑1 -18.81%
- 8 MiMo-7B-RL ↓1 -27.10%
Eng. (G.)
8 models we track have a score here.
- 1 GPT-4o mini ↑3 10%
- 2 MiMo-7B-RL ↑5 -4.41%
- 3
Llama 3.3 70B Instruct -12.53%
- 4
DeepSeek-R1-Distill-Qwen-7B ↑4 -19.59%
- 5
QwQ-32B ↓3 -22.53%
- 6 GLM-4-9B-Chat -25.35%
- 7
Qwen3-32B ↓6 -26.99%
- 8
Qwen3-8B ↓3 -27.13%
Che. (G.)
8 models we track have a score here.
- 1
Qwen3-8B ↑4 59.3%
- 2
QwQ-32B 48.9%
- 3 GPT-4o mini ↑1 36.4%
- 4
Qwen3-32B ↓3 36.3%
- 5 GLM-4-9B-Chat ↑1 7.18%
- 6 MiMo-7B-RL ↑1 1.04%
- 7
Llama 3.3 70B Instruct ↓4 -15.83%
- 8
DeepSeek-R1-Distill-Qwen-7B -44.58%
Number one over time
Lab claims count from the day the card came out, independent runs from the model’s release.
Source: The SAS-Bench paper
Scores over time
All the scores we have, one dot each, so a model tested at several settings appears more than once.
Source: Official SAS-Bench leaderboard
No lab model card reports it yet
We read the headline results of 540 launch posts from 57 labs, and none of them list this benchmark.
More in Law, finance and medicine
- MedXpertQA-MM In use Models answer expert-level medical exam questions that combine images with clinical records and test results. Scoring is accuracy.
Qwen 3.8 Max 80.4%
- HealthBench Professional In use 525 physician-authored tasks from real clinician chats (care consult, documentation, research) with physician-adjudicated rubrics. Claude Sonnet 5.5 69.2%
- HealthBench In use 5,000 realistic health conversations graded against physician-written rubrics. A length-adjusted variant penalizes verbosity. GPT-5 67.2%
- FinSearchComp In use Agents search the web to fetch live financial data, look up history, and run multi-step investigations, across 635 expert-written questions. Scored by answer accuracy.
Ling-3.1-flash 81.7%
- KorMedMCQA In use Models answer 7,469 multiple-choice questions from Korean licensing exams for doctors, nurses, pharmacists and dentists, held from 2012 to 2024. o1-preview 92.7%
- Harvey LAB-AA v1.1 In use Artificial Analysis's run of Harvey's Legal Agent Benchmark on Harvey's newer private set of 120 tasks across 24 practice areas. A deliverable with a material hallucination scores zero, and the headline is the share of judges who find every rubric criterion passed. GPT-6.2 Astra 15%
Data audit
9 of 10 checks pass, 1 warning. Look first at the Newest score check. The newest came 19 months ago, in April 2025, for MiMo-7B-RL. Every benchmark
Catalog record 2 pass
- Pass Task access All tasks are public.
- Pass Confidence Checked against a primary source.
Scores 1 warning 3 pass 1 note
- Pass Independent runs Scores from the benchmark’s paper (16 entries).
- Pass Models scored 8 models we track, plus 8 other entries on its leaderboard.
- Warning Newest score The newest came 19 months ago, in April 2025, for MiMo-7B-RL.
- Note Frontier models None of the top 10 has a score. The strongest tested is Qwen3-32B (#135).
- Pass Headroom The best is 67.2%, under 85%, so it still separates models.
Leaderboard import 3 pass
- Pass Importer We read the board on github.com: 8 rows for models we track and 8 other entries at the last read.
- Pass Last read November 20th, 2026.
- Pass Read worked The last read worked.
Consistency 1 pass 1 note
- Note Lab claims No model has both a lab number and an independent run.
- Pass Known issues No published audit found problems.