AdvancedIF
Chat & Writing
Meta's AdvancedIF tests complex, multi-turn and system-prompt instruction following on 1,645 prompts, each graded against expert-written rubrics by an LLM judge. Pass@1 is the share of responses that satisfy every rubric item of their prompt.
Top models (higher is better)
| Model | Score |
|---|---|
| Gemini 3.1 Pro Preview | 86.7 |
| Gemini 3.6 Flash | 85.3 |
| Claude Fable 5.1 | 84.2 |
| DeepSeek V4.1 Flash | 84.0 |
| Claude Sonnet 5.5 | 83.4 |
| Claude Opus 5.5 | 82.8 |
| GPT-6.1 Sol | 82.4 |
| GPT-5.6 Sol Pro | 82.1 |
| GPT-6 Astra | 82.0 |
| GPT-5.5 | 81.3 |
| GPT-5.4 | 81.0 |
| GPT-6 Sol | 80.9 |
| Grok 4.6 | 80.9 |
| Claude Haiku 5.5 | 80.8 |
| Claude Opus 5 | 80.5 |
| Grok 4.7 | 80.2 |
| Muse Spark 1.1 | 79.4 |
| Grok 4.5 | 79.3 |
| GLM-5.3 | 78.2 |
| GPT-5.6 Terra | 78.1 |
| Qwen3.5 397B A17B | 78.1 |
| Inkling | 77.6 |
| GLM-5.2 | 77.4 |
| GPT-5.6 Luna | 76.0 |
| GPT-6 Luna | 74.9 |
| Kimi K2.7 Code | 73.8 |
| Opus 4.8 | 73.7 |
| Gemini 3.5 Flash | 73.7 |
| Sonnet 5 | 73.3 |
| Opus 4.7 | 73.1 |
| Sonnet 4.6 | 70.8 |
| Kimi K2 Thinking | 68.5 |
| gpt-oss-120b | 68.3 |
| Claude Fable 5 | 65.2 |
| Kimi K3 | 58.7 |
| MiniMax M3 | 56.3 |
| Nemotron 3 Ultra 550B A55B (NVFP4) | 53.7 |
| DeepSeek-V3.2 | 51.3 |
| MiniMax M2.7 | 47.8 |