[mle]/DaBench

Claude Haiku 5.5 vs DeepSeek V4.1 Flash 1017

Side A
Model page
Side B
Model page
Overall
70
vs
Overall
56

Claude Haiku 5.5 is clearly better. It wins 32 of the 42 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. DeepSeek V4.1 Flash 1017 spends 1.6× as many thinking tokens per test case.

Thinking tokensAnswer tokens
TaskAB
Overall
7025.5s
3.9K, 78% thinking5.3K, 91% thinking
5626.0s
Agents
6139.5s
3.7K, 61% thinking3.3K, 71% thinking
5826.7s
Structured Extraction
9013.5s
2.6K, 83% thinking4.2K, 93% thinking
8117.2s
Coding
8630.6s
6.8K, 76% thinking9.4K, 95% thinking
4840.8s
Grounded Answers
7613.9s
2.2K, 78% thinking4.4K, 91% thinking
7119.8s
Instruction Following
9431.5s
4.3K, 82% thinking5.5K, 91% thinking
7625.6s
Business Logic
6726.0s
5.5K, 95% thinking6.6K, 97% thinking
6726.6s
Security & permissions
6723.4s
2.7K, 70% thinking4.7K, 87% thinking
4826.6s
Diagnosis & investigation
1725.5s
2.3K, 50% thinking3.2K, 74% thinking
020.2s

What you'd run

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

SpecAB
Speed
225 tok/s
243 tok/s (better)
Machine
Hosted
Hosted
Memory at 32K
—
—
Context
1M
1M
Cost per run
$0.42 (better)
$0.75
Cost per case
$0.0028 (better)
$0.0049
Time per case
25.5s (better)
26.0s
Quantization
—
—
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.

Claude Haiku 5.5DeepSeek V4.1 Flash 1017
025507510050100150Perfect run7056

Every test case

153 test cases

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

32 won by Claude Haiku 5.510 won by DeepSeek V4.1 Flash 1017111 tied