# benchmarks

Anthropic found that agentic coding benchmark scores can swing by up to 6 percentage points based on hardware configuration alone — often more than the gap between top models on leaderboards. Next time someone claims a 2-3% lead, ask them what VM they ran on.