Benchmark Almanac

BRUMO 2025

Gemini 3 Flash 100% leads since December 2025
Not in model cards

Models answer the 30 final-answer problems of the 2025 Brown University Math Olympiad. MathArena scores the average accuracy over four attempts.

MathArena ran 45 models and later marked the set deprecated. MathArena did not run a 2026 edition.

Released
April 2025
Kind
Competition problems with answers
Level
Competition
Scored
Checked automatically
Answers
Short answer
Tasks
All public
Strongest model tested
GPT-5.2 #50 on the capability index
Last model added
February 2026
Scores refreshed
November 12th, 2026
42
Models scored on MathArena’s leaderboard

2 models share the top score of 100% among models released by February 2026

Best score per model from MathArena’s leaderboard, read on November 12th, 2026. Where a model ran with several settings or harnesses, we keep its best. Every run is in the next chart.

The newest model with a score here came out in February 2026, and nobody has tested a newer one. Among the 20 highest-ranked models on the capability index, these have no score yet:

42 models. Scroll sideways for the rest.

  1. 100.0 Gemini 3 Flash
  2. 100.0 Step 3.5 Flash
  3. 99.2 DeepSeek-V3.2-Speciale
  4. 99.2 GLM-5
  5. 98.3 Gemini 3 Pro
  6. 98.3 GPT-5.2
  7. 98.3 Kimi K2.5
  8. 97.5 Grok 4.1 Fast
  9. 96.7 DeepSeek-V3.2
  10. 95.8 o3
  11. 95.8 Grok 4 Fast
  12. 95.8 DeepSeek-V3.2-Exp
  13. 95.0 Grok 4
  14. 94.2 GLM-4.6
  15. 93.3 Kimi K2 Thinking
  16. 93.3 GPT-5.1
  17. 92.5 DeepSeek-R1
  18. 92.5 GLM-4.5
  19. 92.5 gpt-oss-120b
  20. 91.7 GPT-5
  21. 90.8 Claude Sonnet 4.5
  22. 90.0 Gemini 2.5 Pro
  23. 90.0 GLM-4.5-Air
  24. 90.0 GPT-5 mini
  25. 90.0 DeepSeek-V3.1
  26. 89.2 Gemini 2.5 Pro
  27. 87.5 QED-Nano
  28. 86.7 o4-mini
  29. 86.7 Qwen3-235B-A22B
  30. 86.7 gpt-oss-20b
  31. 85.8 Falcon H1R 7B
  32. 85.0 Grok-3 mini
  33. 83.3 K2-Think
  34. 83.3 Gemini 2.5 Flash
  35. 81.7 Claude Opus 4
  36. 80.8 DeepSeek-R1
  37. 80.8 GPT-5 nano
  38. 77.5 Qwen3-30B-A3B
  39. 68.3 DeepSeek-R1-Distill-Qwen-14B
  40. 68.3 DeepSeek-R1-Distill-Qwen-32B
  41. 66.7 DeepSeek-R1-Distill-Llama-70B
  42. 65.8 Claude 3.7 Sonnet
  1. 1 Gemini 3 Flash high 100%
  2. 2 Step 3.5 Flash 100%
  3. 3 DeepSeek-V3.2-Speciale 99.2%
  4. 4 GLM-5 99.2%
  5. 5 Gemini 3 Pro 98.3%
  6. 6 GPT-5.2 high 98.3%
  7. 7 Kimi K2.5 high 98.3%
  8. 8 Grok 4.1 Fast reasoning 97.5%
  9. 9 DeepSeek-V3.2 thinking 96.7%
  10. 10 o3 high 95.8%
  11. 11 Grok 4 Fast 95.8%
  12. 12 DeepSeek-V3.2-Exp thinking 95.8%
  13. 13 Grok 4 95%
  14. 14 GLM-4.6 94.2%
  15. 15 Kimi K2 Thinking 93.3%
  16. 16 GPT-5.1 high 93.3%
  17. 17 DeepSeek-R1 May 2025 92.5%
  18. 18 GLM-4.5 92.5%
  19. 19 gpt-oss-120b high 92.5%
  20. 20 GPT-5 high 91.7%
  21. 21 Claude Sonnet 4.5 thinking 90.8%
  22. 22 Gemini 2.5 Pro Mar 2025 90%
  23. 23 GLM-4.5-Air 90%
  24. 24 GPT-5 mini high 90%
  25. 25 DeepSeek-V3.1 thinking 90%
  26. 26 Gemini 2.5 Pro May 2025 89.2%
  27. 27 QED-Nano 87.5%
  28. 28 o4-mini high 86.7%
  29. 29 Qwen3-235B-A22B 86.7%
  30. 30 gpt-oss-20b high 86.7%
  31. 31 Falcon H1R 7B 85.8%
  32. 32 Grok-3 mini high 85%
  33. 33 K2-Think 83.3%
  34. 34 Gemini 2.5 Flash Sep 2025, thinking 83.3%
  35. 35 Claude Opus 4 thinking 81.7%
  36. 36 DeepSeek-R1 80.8%
  37. 37 GPT-5 nano high 80.8%
  38. 38 Qwen3-30B-A3B 77.5%
  39. 39 DeepSeek-R1-Distill-Qwen-14B 68.3%
  40. 40 DeepSeek-R1-Distill-Qwen-32B 68.3%
  41. 41 DeepSeek-R1-Distill-Llama-70B 66.7%
  42. 42 Claude 3.7 Sonnet thinking 65.8%

Source: MathArena’s leaderboard

#ModelScoreSettingDateSource
1 Gemini 3 Flash 100% high Dec 2025 Official leaderboard
2 Step 3.5 Flash 100% Feb 2026 Official leaderboard
3 DeepSeek-V3.2-Speciale 99.2% Dec 2025 Official leaderboard
4 GLM-5 99.2% Feb 2026 Official leaderboard
5 Gemini 3 Pro 98.3% Nov 2025 Official leaderboard
6 GPT-5.2 98.3% high Dec 2025 Official leaderboard
7 Kimi K2.5 98.3% high Jan 2026 Official leaderboard
8 Grok 4.1 Fast 97.5% reasoning Nov 2025 Official leaderboard
9 DeepSeek-V3.2 96.7% thinking Dec 2025 Official leaderboard
10 o3 95.8% high Apr 2025 Official leaderboard
11 Grok 4 Fast 95.8% Sep 2025 Official leaderboard
12 DeepSeek-V3.2-Exp 95.8% thinking Sep 2025 Official leaderboard
13 Grok 4 95% Jul 2025 Official leaderboard
14 GLM-4.6 94.2% Sep 2025 Official leaderboard
15 Kimi K2 Thinking 93.3% Nov 2025 Official leaderboard
16 GPT-5.1 93.3% high Nov 2025 Official leaderboard
17 DeepSeek-R1 May 2025 92.5% May 2025 Official leaderboard
18 GLM-4.5 92.5% Aug 2025 Official leaderboard
19 gpt-oss-120b 92.5% high Aug 2025 Official leaderboard
20 GPT-5 91.7% high Aug 2025 Official leaderboard
21 Claude Sonnet 4.5 90.8% thinking Sep 2025 Official leaderboard
22 Gemini 2.5 Pro Mar 2025 90% Mar 2025 Official leaderboard
23 GLM-4.5-Air 90% Jul 2025 Official leaderboard
24 GPT-5 mini 90% high Aug 2025 Official leaderboard
25 DeepSeek-V3.1 90% thinking Aug 2025 Official leaderboard
26 Gemini 2.5 Pro May 2025 89.2% May 2025 Official leaderboard
27 QED-Nano 87.5% Feb 2026 Official leaderboard
28 o4-mini 86.7% high Apr 2025 Official leaderboard
29 Qwen3-235B-A22B 86.7% Apr 2025 Official leaderboard
30 gpt-oss-20b 86.7% high Aug 2025 Official leaderboard
31 Falcon H1R 7B 85.8% Jan 2026 Official leaderboard
32 Grok-3 mini 85% high Apr 2025 Official leaderboard
33 K2-Think 83.3% Sep 2025 Official leaderboard
34 Gemini 2.5 Flash Sep 2025 83.3% thinking Sep 2025 Official leaderboard
35 Claude Opus 4 81.7% thinking May 2025 Official leaderboard
36 DeepSeek-R1 80.8% Jan 2025 Official leaderboard
37 GPT-5 nano 80.8% high Aug 2025 Official leaderboard
38 Qwen3-30B-A3B 77.5% Apr 2025 Official leaderboard
39 DeepSeek-R1-Distill-Qwen-14B 68.3% Jan 2025 Official leaderboard
40 DeepSeek-R1-Distill-Qwen-32B 68.3% Jan 2025 Official leaderboard
41 DeepSeek-R1-Distill-Llama-70B 66.7% Jan 2025 Official leaderboard
42 Claude 3.7 Sonnet 65.8% thinking Feb 2025 Official leaderboard

The best score, 100%, costs $0.008 a task

Runs on the cost frontier score higher than any run that's cheaper. Grok 4 Fast comes within 4.2 points of the best for $0.004 a task, about half the cost. The costs are the leaderboard's own estimates, so compare them only within this benchmark.

DeepSeek xAI StepFun Cost frontier One model at each effort level
0% 20% 40% 60% 80% 100% $0.001 $0.01 $0.10 $1.00 Cost per task, log scale Step 3.5 Flash Grok 4 Fast DeepSeek-R1-Distill-Qwen-14B 0% 20% 40% 60% 80% 100% $0.001 $0.01 $0.10 $1.00 Cost per task, log scale Step 3.5 Flash DeepSeek-V3.2-Speciale

Source: MathArena’s leaderboard

ModelEffortHarnessScoreCost per taskOutput tokens
Gemini 3 Flash high 100% $0.05 15,190
Step 3.5 Flash 100% $0.008 25,178
DeepSeek-V3.2-Speciale 99.2% $0.007 17,472
GLM-5 99.2% $0.07 20,400
Gemini 3 Pro 98.3% $0.15 12,732
GPT-5.2 high 98.3% $0.08 5,989
Kimi K2.5 high 98.3% $0.06 20,245
Grok 4.1 Fast reasoning 97.5% $0.005 8,934
DeepSeek-V3.2 thinking 96.7% $0.005 12,968
o3 high 95.8% $0.08 10,065
Grok 4 Fast 95.8% $0.004 7,517
DeepSeek-V3.2-Exp thinking 95.8% $0.005 10,917
Grok 4 95% $0.16 10,956
GLM-4.6 94.2% $0.03 13,740
Kimi K2 Thinking 93.3% $0.05 19,263
GPT-5.1 high 93.3% $0.17 16,627
DeepSeek-R1 May 2025 92.5% $0.04 18,685
GLM-4.5 92.5% $0.04 19,732
gpt-oss-120b high 92.5% $0.01 18,200
GPT-5 high 91.7% $0.11 10,760
Claude Sonnet 4.5 thinking 90.8% $0.23 15,109
Gemini 2.5 Pro Mar 2025 90% $0.18 17,840
GLM-4.5-Air 90% $0.02 20,348
GPT-5 mini high 90% $0.03 13,545
DeepSeek-V3.1 thinking 90% $0.03 12,375
Gemini 2.5 Pro May 2025 89.2% 3,143
QED-Nano 87.5% 31,439
o4-mini high 86.7% $0.04 7,492
Qwen3-235B-A22B 86.7% $0.007 12,185
gpt-oss-20b high 86.7% $0.007 34,029
Falcon H1R 7B 85.8% $0.004 19,481
Grok-3 mini high 85% $0.007 14,850
o4-mini medium 84.2% $0.02 4,824
K2-Think 83.3%
Gemini 2.5 Flash Sep 2025 thinking 83.3% $0.07 21,389
Claude Opus 4 thinking 81.7% $0.98 12,974
DeepSeek-R1 80.8% $0.02 9,086
GPT-5 nano high 80.8% $0.01 25,061
Qwen3-30B-A3B 77.5% $0.004 14,564
DeepSeek-R1-Distill-Qwen-14B 68.3% $0.002 10,864
DeepSeek-R1-Distill-Qwen-32B 68.3% $0.003 10,884
DeepSeek-R1-Distill-Llama-70B 66.7% $0.006 9,313
o4-mini low 66.7% $0.008 1,874
Claude 3.7 Sonnet thinking 65.8% $0.33 22,001
Grok-3 mini low 65.8% $0.003 5,423

5 models have held number one on MathArena’s leaderboard since its release in April 2025

Lab claims count from the day the card came out, independent runs from the model’s release.

Google DeepMind OpenAI DeepSeek
MathArena leaderboard Passed 85% in Apr 2025, by Gemini 2.5 Pro
Gemini 3 Flash 100%

Source: MathArena’s leaderboard

Gemini 2.5 Pro on the official leaderboard passed 85% in March 2025

All the scores we have, one dot each, so a model tested at several settings appears more than once.

On MathArena’s leaderboard Best so far
0% 20% 40% 60% 80% 100% 2025 2026 Near the ceiling (85%) Gemini 3 Flash o3 Gemini 2.5 Pro DeepSeek-R1 0% 20% 40% 60% 80% 100% 2025 2026 Near the ceiling (85%) Gemini 3 Flash Gemini 2.5 Pro

Source: MathArena’s leaderboard

No lab model card reports it yet

We read the headline results of 522 launch posts from 52 labs, and none of them list this benchmark.

More in Math

All 113 in Math
  • AIME 2026 In use Models solve the 30 integer-answer problems from the 2026 AIME I and II math competitions. Scored by exact answer match. GPT-5.5 100%
  • IMO-AnswerBench In use 400 olympiad problems with short verifiable answers, robustified against memorization. Scored by answer match. GLM-5.2 91%
  • HMMT 2026 In use Models solve the 33 final-answer problems from the February 2026 Harvard-MIT Mathematics Tournament. Scored by exact answer match. GPT-5.5 98.5%
  • PolyMath In use Models solve 9,000 parallel math problems in 18 languages across four difficulty levels, from school math to frontier math. Final answers are checked. Qwen3.7-Max 86.5%
  • MathArena (Apex, IMO, USAMO) In use Evaluates models on newly released math competitions after release. Final-answer or human/LLM-judged proof scores. Claude Opus 4.8 81.3%
  • MathArena Apex Shortlist In use Models solve 47 hard final-answer problems that MathArena selected from 2025 competitions, where the best models scored about 50% at selection. Scored by accuracy over repeated attempts. GPT-5.5 98.4%