[mle]/DaBench

Beam 1.0 501B A23B vs Claude Haiku 5.5

Side A
Model page
Side B
Model page
Overall
57
vs
Overall
70

Claude Haiku 5.5 is clearly better. It wins 39 of the 47 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 1.2× as many thinking tokens per test case.

Thinking tokensAnswer tokens
TaskAB
Overall
5717.2s
4K, 87% thinking3.9K, 78% thinking
7025.5s
Agents
5521.8s
4.8K, 90% thinking3.7K, 61% thinking
6139.5s
Structured Extraction
819.98s
3.4K, 91% thinking2.6K, 83% thinking
9013.5s
Coding
4819.9s
4.5K, 70% thinking6.8K, 76% thinking
8630.6s
Grounded Answers
6911.6s
2.7K, 88% thinking2.2K, 78% thinking
7613.9s
Instruction Following
7820.9s
5.6K, 90% thinking4.3K, 82% thinking
9431.5s
Business Logic
5714.0s
4.2K, 95% thinking5.5K, 95% thinking
6726.0s
Security & permissions
5120.6s
3.2K, 88% thinking2.7K, 70% thinking
6723.4s
Diagnosis & investigation
1724.3s
3.1K, 90% thinking2.3K, 50% thinking
1725.5s

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)
225 tok/s
Machine
Hosted
Hosted
Memory at 32K
—
—
Context
1M
1M
Cost per run
$0.83
$0.42 (better)
Cost per case
$0.0054
$0.0028 (better)
Time per case
17.2s (better)
25.5s
Quantization
FP8
—
Harness
openrouter
openrouter

The race

No lead changes

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

Beam 1.0 501B A23BClaude Haiku 5.5
025507510050100150Perfect run5770

Every test case

153 test cases

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

8 won by Beam 1.0 501B A23B39 won by Claude Haiku 5.5106 tied