Last-Mile Instruction Clarification
Given a customer's delivery note, rewrite it as concise driver steps while preserving access constraints and flagging ambiguous or unsafe instructions.
View token rates
Compare 2 router rates · observed 9/11/2026
Compare 2 router rates · observed 9/11/2026
All 12 model results and methodology
Generated 100/100 examples
| Model | Tier | Quality | Judged | Scenario cost |
|---|---|---|---|---|
| Nemotron Nano 9B v2 | small | 36% | 100/100 | $28.32 |
| Qwen3 235B A22B | mid | 62% | 100/100 | $39.99 |
| DeepSeek V3 | mid | 69% | 100/100 | $19.23 |
| Mistral Large 2407 | mid | 66% | 100/100 | $43.57 |
| Arcee Trinity Large Thinking | mid | 67% | 100/100 | $61.20 |
| GPT-5.4 | frontier | 94% | 100/100 | $248 |
| Claude Opus 4.7 | frontier | 94% | 100/100 | $1107 |
| Gemini 3.1 Pro Preview | frontier | 82% | 100/100 | $1305 |
| Gemma 4 E4B IT | small | 21% | 100/100 | $1.76 |
| Granite 4.1 8B | small | 23% | 100/100 | $3.71 |
| Ministral 8B Instruct 2410 | small | 15% | 100/100 | $5.70 |
| Qwen3 4B Instruct 2507 | small | 27% | 100/100 | $16.24 |
LLM-judge pass rate on 100 synthetic examples. Generator: gpt-5.2. Judge: gpt-5.2. Evaluated 2026-09-09T11:07:02.682Z.
Directional: measured on a synthetic eval set generated by drydock. Cost/latency are not yet captured for taskrouter-run benchmarks.
drydock