AIEvals

Benchmarks

SAGE

Whether a model can grade student mathematics work against rubrics — the assessment load teachers carry — rather than only solving the problem itself.

EducationVals AI90 models run by Vals AIUpdated 29 September 2026

How it is run

The dataset spans automata, linear algebra, calculus and partial differential equations, drawn from course materials and past examination questions, with rubric-based grading. Vals also measures which answers each model grades too harshly or too leniently.

What it cannot tell you

Substantial grading errors remain even in the strongest runs. Vals’ analysis finds correct answers easier to grade than incorrect ones, with grading bias varying by model; the overall score can hide that difference.

Leader
Claude Opus 4.756.10% · Vals AI's own list
Top-ten spread
4.25%Leader minus tenth place, or last place in smaller fields.
Median
44.16%of the 90 rows published
No published result
14of 46 models on this site

Independent leaderboard · Vals AI

Top 90 of 90 runs by Vals AI, one harness.

#ModelScore
1Claude Opus 4.756.10%
2Gemma 4 31b It55.03%
3Claude Opus 4.854.79%
4Kimi K354.26%
5Gemini 4 Argon53.65%
6GPT-5.6 Sol52.56%
7Qwen3.8 27b52.40%
8Claude Opus 4 5 20251101 Thinking52.09%
9Claude Fable 551.89%
10Gemini 3 Flash Preview51.85%
11Claude Sonnet 5.551.77%
12Claude Opus 4 6 Thinking51.58%
13GPT-5.551.53%
14Qwen3.8 Max51.25%
15Gpt 5.4 Mini 2026 03 1750.81%
16MiniMax M350.57%
17Kimi K2.650.22%
18Gemini 3.5 Flash49.88%
19Kimi K2.5 Thinking49.87%
20Gemini 3.1 Flash Lite Preview49.54%
21Claude Opus 549.43%
22Gpt 5.2 2025 12 1149.27%
23Gemini 3.7 Flash49.23%
24Claude Sonnet 548.92%
25Gemini 3.6 Flash48.69%
26Gemini 3.1 Pro48.68%
27Claude Fable 5.148.53%
28GPT-6 Luna48.09%
29DeepSeek V4.1 Flash47.88%
30Muse Spark 1.247.66%
31Gemini 3 Pro Preview47.62%
32Gemini 3.5 Flash Lite47.30%
33GPT-5.6 Terra47.00%
34Claude Sonnet 4 646.58%
35GPT-6.1 Sol46.54%
36GPT-6 Astra46.37%
37Claude Opus 5.545.83%
38Muse Spark 1 145.58%
38Qwen3.6 27b45.58%
40MiMo V2.6 Pro45.05%
41Claude Opus 4 5 2025110145.00%
42Qwen3.6 Plus44.86%
43GPT-6 Sol44.79%
44Gemini 2.5 Flash44.76%
45GPT-5.6 Luna44.22%
46Gemini 2.5 Flash Preview 09 202544.11%
47Gpt 5 2025 08 0743.68%
48MiMo-V2.6-Flash43.53%
49Gpt 5.4 2026 03 0543.31%
50Mimo V2.543.27%
51Gpt 5.1 2025 11 1343.23%
52Gpt 5 Mini 2025 08 0742.99%
53Qwen3.5 Flash42.49%
54Qwen3 Vl Plus 2025 09 2342.10%
55Gemini 2.5 Pro41.92%
56O3 2025 04 1641.77%
57O4 Mini 2025 04 1641.06%
58Grok 4.740.79%
59Gemini 2.5 Flash Thinking39.60%
60Gemini 2.5 Flash Preview 09 2025 Thinking39.38%
61Qwen3.7 Plus39.25%
62Grok 4.20 0309 Reasoning38.24%
63Gpt 5.4 Nano 2026 03 1738.08%
64Mistral Medium 3.537.61%
65Inkling36.55%
66Claude Sonnet 4 5 20250929 Thinking36.06%
67Llama4 Maverick Instruct Basic35.70%
68Gemini 3.8 Flash35.06%
69Claude Sonnet 4 2025051435.00%
70Grok 4.534.97%
71Llama 4 Scout 17B 16E Instruct34.83%
72Claude Sonnet 4 5 2025092932.88%
73Inkling Small32.74%
74Claude Sonnet 4 20250514 Thinking32.01%
75Claude Haiku 4 5 20251001 Thinking31.82%
76Claude Opus 4 1 2025080531.38%
77Grok 4 1 Fast Reasoning31.33%
78Gemini 2.5 Flash Lite Preview 09 2025 Thinking30.78%
79Qwen3.5 Plus Thinking30.40%
80Claude Opus 4 1 20250805 Thinking30.39%
81Gpt 5 Nano 2025 08 0730.38%
82Grok 4 Fast Reasoning29.76%
83Grok 4.628.90%
84Gemini 2.5 Flash Lite Preview 09 202527.97%
85Gemini 2.5 Flash Lite27.53%
86Grok 4 070925.10%
87Mistral Large 251224.59%
88Grok 4.319.74%
89Grok 4 Fast Non Reasoning15.99%
90Grok 4 1 Fast Non Reasoning11.91%
Source: www.vals.ai/benchmarks/sage

No published result tracked

14 of the 46 models on this site have no published number on this benchmark, so their cells in the home table stay blank rather than reading as zero.

Step 5 PreviewGLM 5.3 FlashQwen3.8 2.4T A95BQwen3.8-Flash-NextQwen3.8 Max (0902)GPT-5.4DeepSeek V4.1 Flash 1017GLM-LegalClaude Haiku 5.5BeamDeepSeek V4 ProGLM 5.3Muse Spark 1.3 MaxHunyuan Hy4 Preview

Sources

Every page used for this benchmark, read on 17 October 2026.