MTAC-IFBench - C-ISR (Claude Code): leaderboard
Multi-turn agentic coding instruction following under the Claude Code framework: Constraint-Instruction Success Rate, the share of instructions whose every constraint is satisfied, averaged over turn buckets 1-10. Higher is better. Source: arxiv.org/abs/2609.14992.
Metric: Constraint-Instruction Success Rate (%, all constraints met, average over turn buckets). Source: arxiv.org. Saturation forecast: Around July 2028. 11 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GLM-5.2 | 12.7 | #80 |
| 2 | Gemini 3.1 Pro (Preview) | 11.1 | #64 |
| 3 | GLM-5.1 | 8.9 | #109 |
| 4 | Claude Opus 4.6 | 8.7 | #67 |
| 5 | Kimi K2.6 | 7.8 | #113 |
| 6 | DeepSeek V4 Pro | 6.4 | #112 |
| 7 | DeepSeek V4 Flash | 3.4 | #155 |
| 8 | Seed 2.0 Pro | 2.6 | #122 |
| 9 | Qwen 3.6 Plus | 2 | #124 |
| 10 | Claude Haiku 4.5 | 1.9 | #287 |
| 11 | Qwen 3.5 27B | 1.1 | #192 |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=mtac-ifbench-c-isr-claude-code · How It Works · Data refreshed daily, snapshot 2026-11-05.