[mle]/DaBench

GPT-6 Luna vs Ling 3.1 Flash

Side A
Model page
Side B
Model page
Overall
61
vs
Overall
45

GPT-6 Luna is clearly better. It wins 39 of the 49 test cases where they differ.

Learn to build evals like this in the AI Engineering Academy

Task by task

A score turns orange only when that model wins clearly more of the task's test cases. Under each score, the average time per test case on the machine it ran on. Ling 3.1 Flash spends 4.2× as many thinking tokens per test case.

Thinking tokensAnswer tokens
TaskAB
Overall
6149.1s
1.9K, 73% thinking7.9K, 73% thinking
451m 16s
Agents
561m 18s
1.3K, 67% thinking4.5K, 66% thinking
4155.9s
Structured Extraction
9514.5s
999, 72% thinking8.1K, 69% thinking
711m 11s
Coding
732m 18s
3.6K, 60% thinking11.3K, 89% thinking
241m 42s
Grounded Answers
6715.6s
1.1K, 72% thinking6.7K, 74% thinking
5258.9s
Instruction Following
8432.3s
2K, 74% thinking11.4K, 77% thinking
761m 55s
Business Logic
3743.1s
3.2K, 93% thinking9.7K, 56% thinking
381m 22s
Security & permissions
6224.5s
1K, 68% thinking5.2K, 74% thinking
5456.1s
Diagnosis & investigation
1737.7s
1.1K, 68% thinking3.2K, 72% thinking
042.1s

What you'd run

The build behind each score. Orange marks the faster, smaller, cheaper or longer-context side.

SpecAB
Speed
90 tok/s
112 tok/s (better)
Machine
Hosted
Hosted
Memory at 32K
—
—
Context
1.1M (better)
256K
Cost per run
$0.10
Not measured
Cost per case
$0.00067
Not measured
Time per case
49.1s (better)
1m 16s
Quantization
—
—
Harness
openrouter
openrouter

The race

2 lead changes

Each model's running score as the test cases go by. The dashed line is a perfect run.

GPT-6 LunaLing 3.1 Flash
025507510050100150Perfect run6145

Every test case

153 test cases

Each square is one test case, colored by the model that scored higher.

39 won by GPT-6 Luna10 won by Ling 3.1 Flash104 tied