Head to head

Any two models from the board, on the same tasks and graders.

Side A

Profile

Hosteddeepseek65 tok/sopenrouter

Side B

Profile

Hostedopenai45 tok/sopenai

Overall

Wins 4 tasks, 31 test cases

Overall

Wins 3 tasks, 17 test cases

DeepSeek V4 Flash 0731 leads by 5 points and wins 4 of 9 tasks.

Learn to build evals like this in the AI Engineering Academy

Task by task

DeepSeek V4 Flash 0731 wins 4 of 9

Average score on each task, best build on each side. The higher score is highlighted. Hover a score for the test cases behind it.

TaskADeepSeek V4 Flash 0731BQwen3.8 27BLead

Overall

Tool Calling

Right tool, right args, right sequence?

+20

Structured Output

Valid JSON matching the schema, correct fields?

+11

RAG / Retrieval QA

A correct answer grounded in the given context, with nothing made up?

Even

Context Recall

Finds the one fact buried in a long context?

Even

Coding

Working code that passes hidden unit tests?

+45

Reasoning & Math

Multi-step math and logic with the exact right answer?

Instruction Following

Obeys format, length, and content constraints to the letter?

Classification

Correct label from a fixed set?

Summarization

Compact summary that keeps the facts and invents none?

+10

The race

10 lead changes

Score piled up over the 238 test cases both models ran, as a share of the suite's maximum. The dashed line is a perfect run.

DeepSeek V4 Flash 0731Qwen3.8 27B

Every test case

238 test cases

One square per test case, in suite order. A square takes the color of the model that scored higher on it.

Tool Calling82

Structured Output14

RAG / Retrieval QA33

Context Recall11

Coding120

Reasoning & Math10

Instruction Following13

Classification20

Summarization24

31 won by DeepSeek V4 Flash 073117 won by Qwen3.8 27B190 tied

Where it breaks

238 shared test cases

On the impossible tier, DeepSeek V4 Flash 0731 passes 38% of test cases and Qwen3.8 27B passes 29%.

TierADeepSeek V4 Flash 0731BQwen3.8 27BLead

Baseline

36 test cases

78%

86%

+8 pts

Hard

69 test cases

61%

57%

+4 pts

Impossible

133 test cases

38%

29%

+9 pts

Tiers come from the suite. Baseline is fair, hard is adversarial, and impossible is built so that nothing solves it. Each number is the share of test cases the model fully passed.

Thinking spend

Qwen3.8 27B spends 1.2× as many thinking tokens per test case.

DeepSeek V4 Flash 0731

3.5Ktokens

Per test case, 76% of it thinking

Qwen3.8 27B

3.4Ktokens

Per test case, 98% of it thinking

TaskADeepSeek V4 Flash 0731BQwen3.8 27B

Tool Calling

4.1K52% thinking

4.8K88% thinking

Structured Output

2.8K78% thinking

2.6K99% thinking

RAG / Retrieval QA

2.9K85% thinking

2.5K100% thinking

Context Recall

75787% thinking

50398% thinking

Coding

6.1K86% thinking

6.2K100% thinking

Reasoning & Math

5.9K73% thinking

5.4K100% thinking

Instruction Following

3.4K71% thinking

2.9K100% thinking

Classification

2.7K81% thinking

2.7K100% thinking

Summarization

1.8K81% thinking

1.9K99% thinking

Average output tokens per test case, and how much of it was thinking.

What you'd run

The build behind each score. The better value in each row is highlighted.

SpecADeepSeek V4 Flash 0731BQwen3.8 27B

Speed

65 tok/s

45 tok/s

Memory

—

Context

—

Cost per run

$0.2742

Quantization

—

Harness

openrouter

openai

Faster, smaller, cheaper and more context count as better. Quantization and harness only describe the builds.