MGSM (Vals AI) leaderboard
- Claude Opus 4.5 leads
- 95.2%
- models measured
- 37
- lead over #2
- +0.8 pts
- median
- 90.7%
Leader on MGSM (Vals AI) · read October 22
On October 22, 2026, 42 of our catalog's models had a MGSM (Vals AI) result, and Anthropic Claude Opus 4.5 led with 95.2%, ahead of Anthropic Claude Opus 4.1 at 94.4%.
Grade-school math word problems translated into many languages. Share solved, as run by Vals AI. An archived board.
Role in the index
MGSM (Vals AI) is shown as published and does not feed any of our indexes. Read the methodology
| # | Model | Result | Latency, as printed | Evidence | Coding index | Output $/Mtok |
|---|---|---|---|---|---|---|
| 1 | Claude Opus 4.5Anthropic | 95.2% | 12 s | 44.0 | $25.00 | |
| 2 | Claude Opus 4.1Anthropic | 94.4% | 14 s | Not benchmarked | $75.00 | |
| 3 | Claude Sonnet 4.5Anthropic | 94.3% | 10 s | 39.5 | $15.00 | |
| 4 | GPT-5.2OpenAI | 94.0% | 11 s | 55.9 | $14.00 | |
| 5 | o4 MiniOpenAI | 93.4% | 8.1 s | Not benchmarked | $4.40 | |
| 6 | Gemini 3 Flash PreviewGoogle | 93.3% | 14 s | 43.7 | $3.00 | |
| 7 | GPT-5.1OpenAI | 93.0% | 8.3 s | 40.4 | $10.00 | |
| 8 | GPT-5OpenAI | 92.8% | 20 s | 38.8 | $10.00 | |
| 9 | GPT-5 MiniOpenAI | 92.6% | 13 s | 34.2 | $2.00 | |
| 10 | Llama 4 MaverickMeta Llama | 92.4% | 2.7 s | Not benchmarked | $1.00 | |
| 11 | R1RetiredDeepSeek | 92.3% | 10 s | Not benchmarked | $2.50 | |
| 12 | Claude Haiku 4.5Anthropic | 92.2% | 9.6 s | 35.6 | $5.00 | |
| 13 | gpt-oss-120bOpenAI | 92.0% | 5.8 s | Not benchmarked | $0.50 | |
| 14 | o3OpenAI | 91.8% | 6.8 s | Not benchmarked | $8.00 | |
| 15 | DeepSeek V3 0324DeepSeek | 91.7% | 18 s | Not benchmarked | $1.07 | |
| 16 | o3 MiniOpenAI | 91.3% | 16 s | Not benchmarked | $4.40 | |
| 17 | Grok 3RetiredxAI | 91.3% | 5.8 s | Not benchmarked | $15.00 | |
| 18 | Kimi K2 0711Moonshot AI | 91.0% | 12 s | Not benchmarked | $2.30 | |
| 19 | Mistral Medium 3.1Mistral AI | 90.9% | 5.0 s | Not benchmarked | $2.00 | |
| 20 | GLM 4.5Z.ai | 90.8% | 1m 30s | Not benchmarked | $2.20 | |
| 21 | GPT-4o (2024-08-06)OpenAI | 90.7% | 6.7 s | Not benchmarked | $10.00 | |
| 22 | GPT-4o (2024-11-20)OpenAI | 90.4% | 4.0 s | Not benchmarked | $10.00 | |
| 23 | Kimi K2 ThinkingMoonshot AI | 90.2% | 48 s | Not benchmarked | $2.50 | |
| 24 | GLM 4.6Z.ai | 89.8% | 34 s | 8.8 | $2.20 | |
| 25 | Gemini 2.5 Flash LiteGoogle | 89.5% | 1.5 s | Not benchmarked | $0.40 | |
| 26 | GPT-5 NanoOpenAI | 89.3% | 23 s | Not benchmarked | $0.40 | |
| 27 | o1OpenAI | 89.3% | 11 s | Not benchmarked | $60.00 | |
| 28 | Gemini 1.5 ProRetiredGoogle | 89.2% | 2.8 s | Not benchmarked | $5.00 | |
| 29 | Gemini 2.0 FlashRetiredGoogle | 89.0% | 1.5 s | Not benchmarked | $0.40 | |
| 30 | gpt-oss-20bOpenAI | 89.0% | 5.8 s | Not benchmarked | $0.15 | |
| 31 | GLM 4.7Z.ai | 88.2% | 1m 03s | 55.1 | $2.20 | |
| 32 | Llama 4 ScoutMeta Llama | 88.0% | 3.5 s | Not benchmarked | $0.50 | |
| 33 | MiniMax M2.1MiniMax | 87.8% | 14 s | Not benchmarked | $1.20 | |
| 34 | GPT-4.1 MiniOpenAI | 87.8% | 2.8 s | Not benchmarked | $1.60 | |
| 35 | GPT-4.1OpenAI | 87.7% | 2.2 s | Not benchmarked | $8.00 | |
| 36 | Mistral LargeMistral AI | 87.2% | 8.0 s | Not benchmarked | $6.00 | |
| 37 | Gemini 1.5 FlashRetiredGoogle | 86.6% | 1.4 s | Not benchmarked | $0.30 | |
| 38 | GPT-4o-miniOpenAI | 86.2% | 4.0 s | Not benchmarked | $0.60 | |
| 39 | DeepSeek V3.2DeepSeek | 86.0% | 28 s | 30.0 | $0.88 | |
| 40 | Command ACohere | 85.7% | 8.4 s | Not benchmarked | $10.00 | |
| 41 | Mistral Large 3 2512Mistral AI | 85.4% | 7.6 s | Not benchmarked | $1.50 | |
| 42 | GPT-4.1 NanoOpenAI | 69.3% | 1.5 s | Not benchmarked | $0.40 |
Only models in our catalog are listed, and each appears once with its best published run. 138 rows were read from the source on October 22, 2026, and rows that match no catalog model are left out. Every other column is a figure the board publishes for the same run.