dfbench

dfbench is the held-out benchmark depthfirst maintains to evaluate frontier models on open-ended defensive security work, across three objectives: detection, validation, and differential analysis.

How dfbench works

dfbench: detection

Recall

20% 30% 40% 50% 60% 70% $0.30 $1 $3 $10 $30 $100 dfs-large1 GPT 5.6 Sol GPT 5.6 Luna Opus 5 Opus 5.5 Grok 4.5 Grok 4.7 Kimi K3 Qwen 3.8 GLM 5.2 DeepSeek v4 Flash Gemini 3.6 Flash Mythos 5 Gemini 3.8 Flash Muse Spark 1.3 DeepSeek v4.1 Flash GPT 6 Luna GPT 6 Sol MiMo-v2.6 Flash MiMo-v2.6 Pro

Cost per task

Figure 1: Vulnerability recall across a security scope, against cost per task. The ideal agent sits in the top left corner. Unless otherwise specified, all models are run on our internal harness.

Every dfbench datapoint, by model and thinking level

Model and thinking level
Thinking
dfs-large1
GPT 5.6 Sol xhigh
GPT 5.6 Sol high
GPT 5.6 Sol med
GPT 5.6 Luna xhigh
GPT 5.6 Luna high
GPT 5.6 Luna med
Opus 5 high
Opus 5 med
Opus 5.5 high
Opus 5.5 med
Grok 4.5 high
Grok 4.7 xhigh
Grok 4.7 high
Kimi K3 max
Kimi K3 high
Qwen 3.8 max
GLM 5.2 xhigh
GLM 5.2 high
DeepSeek v4 Flash max
DeepSeek v4 Flash high
Gemini 3.6 Flash high
Mythos 5*
Gemini 3.8 Flash high
Muse Spark 1.3 xhigh
DeepSeek v4.1 Flash high
GPT 6 Luna xhigh
GPT 6 Luna high
GPT 6 Luna med
GPT 6 Sol xhigh
GPT 6 Sol high
GPT 6 Sol med
MiMo-v2.6 Flash
MiMo-v2.6 Pro
Detection cost and overall recall
$6.77 62.2%
$43.37 65.7%
$26.47 59.6%
$9.51 48.9%
$2.53 52.4%
$1.57 39.6%
$0.69 28%
$22.66 45.4%
$21.89 47.8%
$8.42 54.9%
$7.66 48.1%
$7.70 57.2%
$19.94 59.0%
$16.02 56.6%
$9.10 48%
$6.22 46.7%
$7.91 40.8%
$9.72 40.7%
$4.09 38.2%
$1.23 31.8%
$1.03 30.1%
$7.44 20.9%
$99.19 69%
$8.03 44.2%
$10.24 47.4%
$1.69 57.3%
$0.43 45.7%
$0.28 38.1%
$0.15 37.3%
$13.30 59.3%
$6.45 57.7%
$3.43 49.6%
$0.86 53.5%
$1.38 55%
Detection recall by domain
66.5% 42.4% 62.3%
67.1% 67.3% 60.7%
62.8% 62.3% 49.2%
54.2% 38.5% 44.3%
56.8% 45.1% 47.5%
48.4% 23.1% 31.1%
31.6% 15.4% 29.5%
54.8% 34.1% 31.1%
45.8% 57.7% 44.3%
60.9% — 48.2%
57.5% — 35.9%
59.1% 51.6% 0%
68.4% 42.1% 53.9%
68.4% 15.8% 55.1%
31.8% 31.8% 30.4%
48.4% 51.8% 37.9%
40.8% 40.8% 40.8%
34.8% 22.4% 44.3%
42.6% 11.1% 44.3%
27.6% 29.5% 45.9%
30.1% 30.1% 30.1%
19.4% 11.5% 32.8%
70.8% 61.4% 71%
50.6% 17.3% 50.8%
55.5% 32% 39.3%
64.1% 54.5% 48.9%
50% 40.4% 39.3%
43.9% 25% 34.4%
40.6% 26.9% 37.7%
61.9% 51.9% 59%
58.7% 56.9% 55.7%
48.4% 51.9% 50.8%
54.9% 47.9% 55.3%
58.6% 38.9% 59.8%
Detection recall by repository scope
63% 52.6%
66.3% 57.9%
63.3% 10.5%
50.6% 26.3%
53.2% 42.1%
40.2% 31.6%
29.3% 10.5%
47% 26.3%
48.2% 42.1%
57.2% 31.3%
49.7% 31.6%
59.9% 25%
61.4% 20%
57.7% 40%
49.8% 34.8%
48.5% 31.6%
40.8% 40.8%
43.4% 20.9%
42.2% 36.8%
32.9% 26.3%
31.9% 8.3%
20.9% 21.1%
70.4% 50%
44.8% 36.3%
48.6% 31.6%
58.5% 42.1%
46.4% 36.8%
39.4% 21.1%
38.6% 21.1%
59.4% 57.9%
58.9% 42.1%
50.2% 42.1%
54.3% 28.6%
— —
Validation precision
18.3%
18%
22.3%
24.6%
23.2%
24.8%
27.3%
39.4%
40.5%
48.5%
44.2%
29.5%
23.9%
30.5%
30.1%
33.3%
29.6%
28.6%
30.8%
24.8%
26.8%
43.5%
24.5%
41.4%
34.1%
36.4%
21.8%
24.5%
33.3%
15.7%
18.9%
22.6%
24.4%
18.1%
F1 from detect recall and validate precision
28.3%
28.3%
32.5%
32.7%
32.2%
30.5%
27.6%
42.2%
43.8%
51.5%
46.1%
38.9%
34.0%
39.7%
37.0%
38.9%
34.3%
33.6%
34.1%
27.9%
28.4%
28.2%
36.2%
42.8%
39.7%
44.5%
29.5%
29.8%
35.2%
24.8%
28.5%
31.1%
33.5%
27.2%
Differential analysis
$1.34 75.6%
$10.67 75.3%
$5.46 76.3%
$2.32 74.4%
$0.62 71%
$0.30 71.3%
$0.13 65.1%
$7.41 70.3%
$5.48 71.3%
$1.78 79.5%
$1.34 75.3%
$3.38 65.1%
$6.56 74.8%
$6.46 70.3%
$6.34 60.3%
$4.18 60.1%
$2.99 72.7%
$2.04 61.4%
$1.56 58.6%
$0.79 66.8%
$0.55 67.4%
$1.51 61%
— —
$1.57 63.8%
$3.12 72.8%
$0.71 70.1%
$0.12 72.1%
$0.09 72.4%
$0.07 69.5%
$3.90 75.2%
$2.06 76.3%
$1.00 71.6%
$0.08 72.7%
$0.12 76.5%
Differential analysis by NEW, KEPT, and CLOSED
45.8% 95% 86%
46.8% 95% 84.1%
45.6% 94.4% 88.9%
39.2% 93.3% 90.7%
38.9% 89.3% 84.8%
32.5% 92.4% 89.1%
22.2% 79.2% 94%
38.2% 79.8% 92.8%
37.4% 83.8% 92.8%
66.7% 86.8% 85.2%
47.7% 91.5% 86.8%
39.9% 74.5% 80.9%
51.2% 90.5% 82.6%
43.9% 86.3% 80.8%
35.5% 72% 73.4%
31.4% 75.3% 73.5%
33.6% 95% 89.6%
25.5% 85.6% 73%
25.3% 76.4% 74.1%
28.4% 83% 89%
27.1% 88.1% 87%
18.3% 68.9% 96.5%
— — —
32% 66.3% 93.1%
47.1% 85.1% 86.2%
48.5% 67.5% 94.2%
44.4% 89% 82.9%
45.8% 91.2% 80.1%
36.1% 87.4% 85%
58.3% 92.9% 74.5%
61.8% 92% 75%
52.8% 86.3% 75.9%
38.7% 89.9% 89.5%
55.6% 92.8% 81.3%
These results will be updated as new models and harnesses are released. Expand detect to see results stratified by example type. * Run with Claude Security.