Benchmarks / Long Context & Instruction
LiveBench Instruction Following
Contamination-limited monthly-refreshed benchmark; category mean over its tasks.
maintained by LiveBench · livebench.ai · 40 observations on record
Versions
1 version
| Version | Tasks |
|---|---|
| Release 2026-06-25 current livebench-if-2026-06-25 | — |
Current leaderboard
Release 2026-06-25 · Accuracy · top 25 of 40 systems
| # | System | Accuracy % |
|---|---|---|
| 1 | Gemini 3.1 Pro Previewhigh effort Google DeepMind | 79.1% |
| 2 | Claude Fable 5max effort Anthropic | 75.8% |
| 3 | Gemini 3.5 Flashhigh effort Google DeepMind | 75.6% |
| 4 | Qwen3 7 Max Alibaba (Qwen) | 74.0% |
| 5 | Nemotron 3 Ultra 550B A55b NVIDIA | 73.5% |
| 6 | Claude Opus 4.8xhigh effort Anthropic | 72.4% |
| 7 | Claude Opus 4.8max effort Anthropic | 72.0% |
| 8 | Claude Fable 5xhigh effort Anthropic | 72.0% |
| 9 | GPT 5.6 Solmax effort OpenAI | 71.8% |
| 10 | Grok 4.5 xAI | 71.5% |
| 11 | GPT 5.5 (2026-04-22)high effort OpenAI | 71.4% |
| 12 | GPT 5.5 (2026-04-22)xhigh effort OpenAI | 70.7% |
| 13 | GPT 5.4 (2026-03-05)xhigh effort OpenAI | 70.2% |
| 14 | Muse Spark 1.1high effort Meta AI | 70.1% |
| 15 | Muse Spark 1.1xhigh effort Meta AI | 69.6% |
| 16 | GPT 5.6 Solxhigh effort OpenAI | 67.4% |
| 17 | GPT 5.4 Nano (2026-03-17)xhigh effort OpenAI | 67.2% |
| 18 | Claude Opus 4.7xhigh effort Anthropic | 66.7% |
| 19 | GPT 5.2 Codex OpenAI | 66.5% |
| 20 | GPT 5.5 (2026-04-22) OpenAI | 65.7% |
| 21 | Grok Build 0.1 xAI | 65.2% |
| 22 | GPT 5.6 Terramax effort OpenAI | 64.6% |
| 23 | Kimi K2 6thinking Moonshot AI | 64.4% |
| 24 | Claude Sonnet 5xhigh effort Anthropic | 63.9% |
| 25 | Claude Opus 4.6high effort Anthropic | 63.3% |
One row per evaluated system — reasoning-effort variants rank separately.