Instruction 基准测试
Compare top models across the benchmark suite that best represents instruction performance. Use this page as the fastest way to inspect the relevant tests, then jump into the full matrix when you want broader context.
2
类别中的基准测试
46
有覆盖的模型
0
有人类基准的基准测试
0
饱和的基准测试
包含内容
The current benchmark set in this category, with context on what each test captures.
Instruction Following Evaluation
Measures instruction-following precision, critical for production applications. Models that score well here are more reliable in structured tasks.
Multi-Turn Benchmark
Tests real conversational ability across turns, not just single-shot performance. Important for chat applications.
Instruction Following Evaluation
Tests whether models follow explicit, verifiable constraints like 'write in more than 400 words' or 'mention AI at least 3 times'. All instructions have objectively verifiable criteria.
Why it matters
Measures instruction-following precision, critical for production applications. Models that score well here are more reliable in structured tasks.
IFEval Scores (46 models)
| # | Model | Score |
|---|---|---|
| 1 | 🥇Gemini 3 Pro | 93.5% |
| 2 | 🥈GPT-5.4 | 93.5% |
| 3 | 🥉GPT-5.2 | 93.0% |
| 4 | Claude Opus 4.6 | 92.8% |
| 5 | GPT-5.1 | 92.5% |
| 6 | Claude 3.7 Sonnet | 92.3% |
| 7 | Llama 3.3 70B | 92.1% |
| 8 | o3 | 92.0% |
| 9 | GPT-5 | 92.0% |
| 10 | Claude Opus 4.5 | 91.5% |
| 11 | Claude Sonnet 4.6 | 91.0% |
| 12 | Grok 4 | 91.0% |
| 13 | Gemini 3 Flash | 91.0% |
| 14 | Claude Sonnet 4 | 90.8% |
| 15 | Claude Opus 4 | 90.5% |
| 16 | o3-mini | 90.2% |
| 17 | Claude Sonnet 4.5 | 90.2% |
| 18 | o4-mini | 90.0% |
| 19 | DeepSeek V3 (March 2025) | 89.0% |
| 20 | GPT-4.1 | 88.2% |
| 21 | Claude 3.5 Sonnet | 88.1% |
| 22 | Llama 4 Maverick | 88.0% |
| 23 | Grok 3 | 88.0% |
| 24 | Llama 3.1 405B | 87.5% |
| 25 | Gemini 2.5 Pro | 87.2% |
| 26 | DeepSeek V3 | 87.1% |
| 27 | o1 | 86.5% |
| 28 | Mistral Large 2 | 86.5% |
| 29 | Gemini 2.5 Flash | 85.5% |
| 30 | DeepSeek R1-0528 | 85.5% |
| 31 | GPT-4o | 84.3% |
| 32 | Claude Haiku 4.5 | 84.0% |
| 33 | Llama 3.1 70B | 83.6% |
| 34 | Qwen 2.5 72B | 83.5% |
| 35 | DeepSeek R1 | 83.3% |
| 36 | Mistral Large 2 | 82.4% |
| 37 | Gemini 2.0 Flash | 82.0% |
| 38 | GPT-4o mini | 80.4% |
| 39 | Phi-4 | 80.1% |
| 40 | Command R7B (12-2024) | 77.1% |
| 41 | Command R+ | 76.6% |
| 42 | Qwen2.5 7B Instruct | 75.8% |
| 43 | Qwen 2.5 Coder 32B | 72.7% |
| 44 | Llama 3.1 8B Instruct | 72.0% |
| 45 | Llama 3.2 3B Instruct | 68.5% |
| 46 | Runway Gen-3 Alpha | 28.0% |
How to Read This Page
Performance Tiers
Model Types
Saturated benchmarks have top models clustered above 90%, making them less useful for comparison.
Scores sourced from official model cards, technical reports, and third-party evaluations (Artificial Analysis, LMSYS Arena). Last updated: 2026-08-08T06:30:09.792Z. Some scores are approximate.
AI benchmarks are grouped into categories like coding, math, reasoning, knowledge, and safety. Each category contains multiple standardized tests that measure specific aspects of model performance. This page focuses on one category so you can compare models within a specific skill area.
Each benchmark has its own scoring method - accuracy percentage, pass rate, Elo rating, or normalized score. We display raw scores from official evaluations and community-run tests. Scores are updated hourly as new evaluation results become available.
A saturated benchmark is one where top models score near the maximum (typically above 95%). This means the benchmark no longer effectively differentiates between the best models, and newer, harder benchmarks are needed to measure progress.