[mle]/DaBench

Beam 1.0 501B A23B vs GPT-6 Luna

Side A
Model page
Side B
Model page
Overall
57
vs
Overall
61

GPT-6 Luna is clearly better. It wins 30 of the 46 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. Beam 1.0 501B A23B spends 2.6× as many thinking tokens per test case.

Thinking tokensAnswer tokens
TaskAB
Overall
5717.2s
4K, 87% thinking1.9K, 73% thinking
6149.1s
Agents
5521.8s
4.8K, 90% thinking1.3K, 67% thinking
561m 18s
Structured Extraction
819.98s
3.4K, 91% thinking999, 72% thinking
9514.5s
Coding
4819.9s
4.5K, 70% thinking3.6K, 60% thinking
732m 18s
Grounded Answers
6911.6s
2.7K, 88% thinking1.1K, 72% thinking
6715.6s
Instruction Following
7820.9s
5.6K, 90% thinking2K, 74% thinking
8432.3s
Business Logic
5714.0s
4.2K, 95% thinking3.2K, 93% thinking
3743.1s
Security & permissions
5120.6s
3.2K, 88% thinking1K, 68% thinking
6224.5s
Diagnosis & investigation
1724.3s
3.1K, 90% thinking1.1K, 68% thinking
1737.7s

What you'd run

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

SpecAB
Speed
274 tok/s (better)
90 tok/s
Machine
Hosted
Hosted
Memory at 32K
—
—
Context
1M
1.1M (better)
Cost per run
$0.83
$0.10 (better)
Cost per case
$0.0054
$0.00067 (better)
Time per case
17.2s (better)
49.1s
Quantization
FP8
—
Harness
openrouter
openrouter

The race

4 lead changes

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

Beam 1.0 501B A23BGPT-6 Luna
025507510050100150Perfect run5761

Every test case

153 test cases

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

16 won by Beam 1.0 501B A23B30 won by GPT-6 Luna107 tied