Ivo-micro1 ContractBench — Completion Rate
Professional Work · 2026-10-13
Share of a model's 630 ContractBench rollouts (three per each of 210 distinct inputs) that end with a gradeable redline. The rest hit the per-turn output limit or fail the validity gate that rejects edits outside tracked changes. ContractBench's headline score counts completed rollouts only, so this rate puts it in context. A diagnostic of output-budget reliability, not of redlining quality.
Top models (higher is better)
| Model | Score |
|---|---|
| GPT-6 Astra | 100.0 |
| Kimi K3 | 100.0 |
| Gemini 3.8 Flash | 99.8 |
| DeepSeek-V4-Pro | 99.5 |
| Grok 4.7 | 99.5 |
| GLM-5.3 | 95.2 |
| Claude Opus 5.5 | 80.8 |