Skip to main content

Vals MGSM (MGSM)

A multilingual benchmark for mathematical questions.

Data verified

How BenchLM shows MGSM

BenchLM mirrors the public Vals AI MGSM leaderboard captured from https://www.vals.ai/benchmarks/mgsm and updated by Vals on January 9, 2026. The snapshot preserves overall scores, uncertainty, latency, cost-per-test metadata, and task-level scores where Vals publishes them.

MGSM is display only on BenchLM. Vals proprietary or Vals-hosted aggregate views are useful context, but BenchLM does not use them as weighted ranking inputs or as a replacement for benchmark-native source records.

75 Vals rows12 task viewspublic datasetTasks: Overall, English, Japanese, Chinese, BengaliDisplay only

MGSM score on MGSM — January 9, 2026

BenchLM mirrors the published mgsm score view for MGSM. Claude Opus 4.5 20251101 Thinking leads the public snapshot at 95.20% , followed by Claude Opus 4.5 (94.76%) and Claude Opus 4.1 20250805 Thinking (94.44%). BenchLM does not use these results to rank models overall.

75 modelsExternal benchmark mirrorsCurrentDisplay onlyUpdated January 9, 2026

The published MGSM snapshot is tightly clustered at the top: Claude Opus 4.5 20251101 Thinking sits at 95.20%, while the third row is only 0.76 points behind. The broader top-10 spread is 1.89 points, so many of the published scores sit in a relatively narrow band.

75 models have been evaluated on MGSM. The benchmark falls in the External benchmark mirrors category. BenchLM tracks this category separately from its weighted global scoring system, so these results are best compared on the dedicated Korean benchmark views. MGSM is currently displayed for reference but excluded from the scoring formula, so it does not directly affect overall rankings.

About MGSM

Year

2026

Tasks

Multilingual grade-school math questions

Format

Accuracy score

Difficulty

Multilingual mathematical reasoning

BenchLM mirrors the public Vals AI MGSM leaderboard as display-only external evidence. The captured snapshot preserves overall scores, task-level scores where Vals publishes them, uncertainty, latency, and cost-per-test metadata. It is excluded from BenchLM weighted rankings.

BenchLM freshness & provenance

Version

MGSM 2026

Refresh cadence

Quarterly

Staleness state

Current

Question availability

Public benchmark set

CurrentDisplay only

BenchLM uses freshness metadata to decide whether a benchmark should still be treated as a strong differentiator, a benchmark to watch, or a display-only reference. For the full scoring policy, see the BenchLM methodology page.

MGSM score table (75 models)

1
Claude Opus 4.5 20251101 Thinkinganthropic/claude-opus-4-5-20251101-thinking
95.20%
2
Claude Opus 4.5anthropic/claude-opus-4-5-20251101
94.76%
3
Claude Opus 4.1 20250805 Thinkinganthropic/claude-opus-4-1-20250805-thinking
94.44%
4
Claude Sonnet 4.5 20250929 Thinkinganthropic/claude-sonnet-4-5-20250929-thinking
94.33%
5
Claude Opus 4.1anthropic/claude-opus-4-1-20250805
94.22%
6
GPT-5.2openai/gpt-5.2-2025-12-11
94.00%
7
Gemini 3 Pro Previewgoogle/gemini-3-pro-preview
93.93%
8
Claude Opus 4anthropic/claude-opus-4-20250514
93.78%
9
O4 Miniopenai/o4-mini-2025-04-16
93.42%
10
Gemini 3 Flash Previewgoogle/gemini-3-flash-preview
93.31%
11
Claude Sonnet 4anthropic/claude-sonnet-4-20250514
93.02%
12
GPT-5.1openai/gpt-5.1-2025-11-13
92.98%
13
Claude 3.7 Sonnet 20250219 Thinkinganthropic/claude-3-7-sonnet-20250219-thinking
92.98%
14
GPT-5openai/gpt-5-2025-08-07
92.84%
15
Claude 3.5 Sonnetanthropic/claude-3-5-sonnet-20241022
92.58%
16
GPT-5 Miniopenai/gpt-5-mini-2025-08-07
92.58%
17
Qwen3 235b A22bfireworks/qwen3-235b-a22b
92.47%
18
Llama4 Maverick Instruct Basicfireworks/llama4-maverick-instruct-basic
92.44%
19
Claude 3.7 Sonnetanthropic/claude-3-7-sonnet-20250219
92.40%
20
DeepSeek R1fireworks/deepseek-r1
92.25%
21
Claude Haiku 4.5 20251001 Thinkinganthropic/claude-haiku-4-5-20251001-thinking
92.15%
22
DeepSeek V3fireworks/deepseek-v3
92.15%
23
Qwen3 Max Previewalibaba/qwen3-max-preview
92.15%
24
GPT Oss 120bfireworks/gpt-oss-120b
92.04%
25
Qwen3 Maxalibaba/qwen3-max
91.82%
26
O3openai/o3-2025-04-16
91.75%
27
DeepSeek V3 0324fireworks/deepseek-v3-0324
91.67%
28
Grok 3grok/grok-3
91.35%
29
O3 Miniopenai/o3-mini-2025-01-31
91.35%
30
Meta Llama Llama 3.3 70B Instruct Turbotogether/meta-llama/Llama-3.3-70B-Instruct-Turbo
91.09%
31
Moonshotai Kimi K2 Instructtogether/moonshotai/Kimi-K2-Instruct
90.95%
32
Grok 4 0709grok/grok-4-0709
90.91%
33
Grok 4 Fast Reasoninggrok/grok-4-fast-reasoning
90.87%
34
Mistral Medium 2505mistralai/mistral-medium-2505
90.87%
35
DeepSeek V3p2fireworks/deepseek-v3p2
90.87%
36
Claude Sonnet 4 20250514 Thinkinganthropic/claude-sonnet-4-20250514-thinking
90.87%
37
GLM 4.5zai/glm-4.5
90.84%
38
GPT-4oopenai/gpt-4o-2024-08-06
90.69%
39
Grok 3 Mini Fast High Reasoninggrok/grok-3-mini-fast-high-reasoning
90.44%
40
GPT-4oopenai/gpt-4o-2024-11-20
90.36%
41
Grok 3 Mini Fast Low Reasoninggrok/grok-3-mini-fast-low-reasoning
90.36%
42
Kimi K2 Thinkingkimi/kimi-k2-thinking
90.15%
43
Gemini 2.5 Flash Preview 09 2025google/gemini-2.5-flash-preview-09-2025
89.85%
44
Gemini 2.5 Flash Preview 09 2025 Thinkinggoogle/gemini-2.5-flash-preview-09-2025-thinking
89.82%
45
GLM 4.6zai/glm-4.6
89.75%
46
Gemini 2.5 Flash Lite Preview 09 2025google/gemini-2.5-flash-lite-preview-09-2025
89.53%
47
Grok 4.1 Fast Reasoninggrok/grok-4-1-fast-reasoning
89.53%
48
O1openai/o1-2024-12-17
89.31%
49
GPT-5 Nanoopenai/gpt-5-nano-2025-08-07
89.31%
50
Gemini 1.5 Pro 002google/gemini-1.5-pro-002
89.20%
51
Gemini 2.0 Flash 001google/gemini-2.0-flash-001
89.02%
52
GPT Oss 20bfireworks/gpt-oss-20b
89.02%
53
Gemini 2.5 Flash Lite Preview 09 2025 Thinkinggoogle/gemini-2.5-flash-lite-preview-09-2025-thinking
88.40%
54
GLM 4.7zai/glm-4.7
88.18%
55
Grok 4 Fast Non Reasoninggrok/grok-4-fast-non-reasoning
88.00%
56
Meta Llama Llama 4 Scout 17B 16E Instructtogether/meta-llama/Llama-4-Scout-17B-16E-Instruct
87.96%
57
MiniMax M2.1minimax/MiniMax-M2.1
87.85%
58
GPT-4.1 Miniopenai/gpt-4.1-mini-2025-04-14
87.78%
59
GPT-4.1openai/gpt-4.1-2025-04-14
87.67%
60
Grok 4.1 Fast Non Reasoninggrok/grok-4-1-fast-non-reasoning
87.56%
61
Mistral Large 2411mistralai/mistral-large-2411
87.24%
62
Gemini 1.5 Flash 002google/gemini-1.5-flash-002
86.58%
63
Magistral Small 2509mistralai/magistral-small-2509
86.25%
64
GPT-4o Miniopenai/gpt-4o-mini-2024-07-18
86.18%
65
Grok 2 1212grok/grok-2-1212
86.15%
66
DeepSeek V3p2 Thinkingfireworks/deepseek-v3p2-thinking
86.04%
67
Command A 03 2025cohere/command-a-03-2025
85.71%
68
Mistral Large 2512mistralai/mistral-large-2512
85.42%
69
Claude 3.5 Haikuanthropic/claude-3-5-haiku-20241022
84.62%
70
Mistral Small 2503mistralai/mistral-small-2503
84.22%
71
Mistral Small 2402mistralai/mistral-small-2402
83.96%
72
Magistral Medium 2509mistralai/magistral-medium-2509
74.62%
73
Jamba Large 1.6ai21labs/jamba-large-1.6
71.24%
74
GPT-4.1 Nanoopenai/gpt-4.1-nano-2025-04-14
69.27%
75
Jamba Mini 1.6ai21labs/jamba-mini-1.6
41.71%

FAQ

What does MGSM measure?

A multilingual benchmark for mathematical questions.

Which model leads the published MGSM snapshot?

Claude Opus 4.5 20251101 Thinking currently leads the published MGSM snapshot with 95.20% mgsm score. BenchLM shows this benchmark for display only and does not use it in overall rankings.

How many models are evaluated on MGSM?

75 AI models are included in BenchLM's mirrored MGSM snapshot, based on the public leaderboard captured on January 9, 2026.

Last updated: January 9, 2026 · mirrored from the public benchmark leaderboard

The AI models change fast. We track them for you.

For engineers, researchers, and the plain curious — a weekly brief on new models, ranking shifts, and pricing changes.

Free. No spam. Unsubscribe anytime.