DAX LLM Benchmark

Which LLM is best at DAX?
DAXBench tests how models understand, write, and reason about DAX.
Developed by Maxim Anatsko in collaboration with SQLBI.

Last updated: Oct 13, 2026

41 models · 94 tasks · DAX Benchmark v2.0

GPT-6 Astra leads, solving 91 of 94 tasks.

  1. 1OpenAI
    GPT-6 AstraAuto
    95.91%91 / 94 solved
  2. 2Anthropic
    Claude Opus 5.5Auto
    95.38%90 / 94 solved
  3. 3OpenAI
    GPT-6.1 SolAuto
    91.42%87 / 94 solved

Model Leaderboard

Ranked by score

ModelTasks
90–100% solved
1
GPT-6 AstraAuto
OpenAI
95.91%
96.81%98.94%91/94
2
Claude Opus 5.5Auto
Anthropic
95.38%
95.74%97.87%90/94
3
GPT-6.1 SolAuto
OpenAI
91.42%
92.55%96.81%87/94
4
Grok 4.7Auto
SpaceXAI
91.22%
92.55%100%87/94
5
Claude Fable 5.1Auto
Anthropic
90.66%
91.49%97.87%86/94
80–89% solved
6
GPT-5.5Auto
OpenAI
86.42%
88.3%95.74%83/94
7
Gemini 3.8 FlashAuto
Google
85.63%
87.23%100%82/94
8
Claude Opus 5Auto
Anthropic
85.4%
86.17%97.87%81/94
9
Claude Sonnet 5.5Auto
Anthropic
84.59%
86.17%97.87%81/94
10
Grok 4.6Auto
SpaceXAI
83.91%
86.17%93.62%81/94
11
GPT-6 LunaAuto
OpenAI
83%
85.11%93.62%80/94
12
Muse Spark 1.3Auto
Meta
82.57%
85.11%95.74%80/94
13
Muse WonderAuto
Meta
82.19%
84.04%92.55%79/94
14
Qwen3.8 Max (0902)Auto
Qwen
81.67%
82.98%97.87%78/94
15
Gemini 3.1 Pro PreviewAuto
Google
81.21%
82.98%96.81%78/94
16
MiMo-V2.6-ProAuto
Xiaomi1 looped
80.73%
81.91%91.49%77/94
17
Claude Haiku 5.5Auto
Anthropic
79.37%
81.91%97.87%77/94
70–79% solved
18
Kimi K3Auto
Moonshot AI
78.05%
79.79%95.74%75/94
19
Kimi K2.7 CodeAuto
Moonshot AI
77.85%
79.79%98.94%75/94
20
GLM 5.3 FlashAuto
Z.AI3 truncated3 looped
75.85%
78.72%91.49%74/94
21
DeepSeek V4.1 FlashAuto
DeepSeek1 truncated5 looped
72.83%
75.53%91.49%71/94
22
Ling 3.1 FlashAuto
inclusionAI4 looped
72.74%
74.47%90.43%70/94
23
Claude Sonnet 5Auto
Anthropic
70.38%
72.34%96.81%68/94
60–69% solved
24
DeepSeek V4 Pro 0813Auto
DeepSeek6 truncated
66.3%
69.15%86.17%65/94
25
GLM 5Auto
Z.AI1 truncated7 looped
66.03%
69.15%89.36%65/94
26
MiniMax M3Auto
Minimax
64.21%
67.02%90.43%63/94
27
Mistral Large 4Auto
Mistral1 truncated5 looped
63.16%
65.96%87.23%62/94
28
GPT-5 MiniAuto
OpenAI
62.96%
65.96%90.43%62/94
29
Muse Glimmer 30BAuto
Meta2 truncated
60.02%
63.83%90.43%60/94
30
MiMo-V2.6-FlashAuto
Xiaomi
58.93%
62.77%93.62%59/94
31
Gemma 4 26B A4B Auto
Google16 looped
56.92%
60.64%80.85%57/94
50–59% solved
32
gpt-oss-120bAuto
OpenAI
54.96%
58.51%92.55%55/94
33
GPT-5.4 MiniAuto
OpenAI
54.41%
57.45%94.68%54/94
34
Nemotron 3 SuperAuto
Nvidia
48.52%
53.19%78.72%50/94
35
Qwen3.8 27BLOW
Qwen
47.62%
51.06%85.11%48/94
36
Qwen3.8 27BAuto
Qwen5 looped
46.63%
51.06%75.53%48/94
40–49% solved
37
Nemotron 3.5 LightningAuto
Nvidia
44.01%
47.87%67.02%45/94
38
Laguna S 2.1Auto
Poolside1 looped
37.63%
40.43%80.85%38/94
30–39% solved
39
gpt-oss-20bAuto
OpenAI1 looped
34.4%
39.36%82.98%37/94
40
Laguna XS 2.1Auto
Poolside
27.52%
31.91%76.6%30/94
20–29% solved
41
Ministral 3 14B 2512Auto
Mistral
23.23%
26.6%59.57%25/94

About This Benchmark

Evaluation Method

Models are tested against DAX tasks of varying complexity using the Contoso sample dataset. Responses are evaluated for syntax correctness and output accuracy.

Scoring System

Harder tasks are worth more points. Correct solutions also earn bonus points for following DAX best practices, writing efficient code, and producing clear, readable output.

Task Categories

Tasks cover aggregation, time intelligence, filtering, calculation, table manipulation, iterator, context transition across basic, intermediate, and advanced levels.

Browse by Category

Evaluating language model capabilities in DAX code generation

Developed by Maxim Anatsko in collaboration with SQLBI

maxanatsko