[mle]/DaBench

DeepSeek V4.1 Flash vs Claude Haiku 5.5

Side A
Model page
Side B
Model page
Overall
64
vs
Overall
70

Claude Haiku 5.5 is clearly better. It wins 21 of the 31 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. The time under each score is the average per test case, on the machine it ran on. DeepSeek V4.1 Flash spends 1.8× as many thinking tokens per test case.

Thinking tokensAnswer tokens
TaskAB
Overall
643m 35s
6K, 91% thinking3.9K, 78% thinking
7025.5s
Agents
464m 42s
6.1K, 82% thinking3.7K, 61% thinking
6139.5s
Structured Extraction
862m 8s
4.4K, 94% thinking2.6K, 83% thinking
9013.5s
Coding
514m 55s
9.7K, 95% thinking6.8K, 76% thinking
8630.6s
Grounded Answers
762m 17s
3.9K, 90% thinking2.2K, 78% thinking
7613.9s
Instruction Following
894m 24s
7.5K, 94% thinking4.3K, 82% thinking
9431.5s
Business Logic
814m 1s
7K, 97% thinking5.5K, 95% thinking
6726.0s
Security & permissions
693m 3s
4.4K, 81% thinking2.7K, 70% thinking
6723.4s
Diagnosis & investigation
172m 8s
3.2K, 67% 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
30 tok/s
225 tok/s (better)
Machine
Hosted
Hosted
Memory at 32K
—
—
Context
1M
1M
Cost per run
$0.39 (better)
$0.42
Cost per case
$0.0026 (better)
$0.0028
Time per case
3m 35s
25.5s (better)
Quantization
FP8
—
Harness
openrouter
openrouter

The race

1 lead change

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

DeepSeek V4.1 FlashClaude Haiku 5.5
0255075100501001506470

Every test case

153 test cases

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

10 won by DeepSeek V4.1 Flash21 won by Claude Haiku 5.5122 tied