Extraction
PHI Exposure Detection in Audit Logs
Given synthetic application logs, identify lines containing possible protected health information, label the PHI type, and avoid flagging safe operational identifiers.
Legal & CompliancePer log
Not benchmarked
No result on this tier
Token profile pending
— /100K
Run this tier to add it
—
Gemma 4 E4B IT
E4B · Off-the-shelf
657 in · 422 out · 6.78s p50
$5.84 /100K
Current price · TrustedRouter
View token rates · observed 9/11/2026
TrustedRouter · $0.0211 input · $0.1055 output /1M
57%
DeepSeek V3
Hosted · Off-the-shelf
550 in · 537 out · 4.06s p50
$68.99 /100K
Best current price · TrustedRouter
Compare 2 router rates · observed 9/11/2026
TrustedRouter · $0.3376 input · $0.9390 output /1M
OpenRouter · $0.2900 input · $1.14 output /1M
61%
GPT-5.4
Hosted · Off-the-shelf
531 in · 345 out · 2.75s p50
$650 /100K
Best current price · OpenRouter
Compare 2 router rates · observed 9/11/2026
OpenRouter · $2.50 input · $15.00 output /1M
TrustedRouter · $2.64 input · $15.82 output /1M
78%
Best accuracy
All 12 model results and methodology
Generated 100/100 examples
| Model | Tier | Quality | Judged | Scenario cost |
|---|---|---|---|---|
| Nemotron Nano 9B v2 | small | 4% | 100/100 | $38.50 |
| Qwen3 235B A22B | mid | 49% | 100/100 | $104 |
| DeepSeek V3 | mid | 61% | 100/100 | $68.99 |
| Mistral Large 2407 | mid | 56% | 100/100 | $126 |
| Arcee Trinity Large Thinking | mid | 8% | 100/100 | $167 |
| GPT-5.4 | frontier | 78% | 100/100 | $650 |
| Claude Opus 4.7 | frontier | 67% | 100/100 | $2281 |
| Gemini 3.1 Pro Preview | frontier | 24% | 100/100 | $1592 |
| Gemma 4 E4B IT | small | 57% | 100/100 | $5.84 |
| Granite 4.1 8B | small | 28% | 100/100 | $5.16 |
| Ministral 8B Instruct 2410 | small | 8% | 100/100 | $17.85 |
| Qwen3 4B Instruct 2507 | small | 22% | 100/100 | $33.03 |
LLM-judge pass rate on 100 synthetic examples. Generator: gpt-5.2. Judge: gpt-5.2. Evaluated 2026-09-08T19:37:02.877Z.
Directional: measured on a synthetic eval set generated by drydock. Cost/latency are not yet captured for taskrouter-run benchmarks.
drydock