PR-Summarizer
Summarizes a pull request diff into a structured review: what changed, why it matters, and risks.
View token rates · observed 9/11/2026
Compare 2 router rates · observed 9/11/2026
Compare 2 router rates · observed 9/11/2026
All 12 model results and methodology
Generated 98/100 examples — partial set accepted at the success threshold
| Model | Tier | Quality | Judged | Scenario cost |
|---|---|---|---|---|
| Nemotron Nano 9B v2 | small | 29% | 98/98 | $24.64 |
| Qwen3 235B A22B | mid | 27% | 98/98 | $122 |
| DeepSeek V3 | mid | 87% | 98/98 | $59.46 |
| Mistral Large 2407 | mid | 92% | 98/98 | $115 |
| Arcee Trinity Large Thinking | mid | 66% | 98/98 | $96.56 |
| GPT-5.4 | frontier | 100% | 98/98 | $806 |
| Claude Opus 4.7 | frontier | 67% | 98/98 | $3539 |
| Gemini 3.1 Pro Preview | frontier | 100% | 98/98 | $1848 |
| Gemma 4 E4B IT | small | 87% | 98/98 | $5.42 |
| Granite 4.1 8B | small | 52% | 98/98 | $9.65 |
| Ministral 8B Instruct 2410 | small | 53% | 98/98 | $17.55 |
| Qwen3 4B Instruct 2507 | small | 19% | 98/98 | $41.10 |
LLM-judge pass rate on 98 synthetic examples. Generator: gpt-5.2. Judge: gpt-5.2. Evaluated 2026-09-06T16:07:02.521Z.
Directional: measured on a synthetic eval set generated by drydock. Cost/latency are not yet captured for taskrouter-run benchmarks.
drydock