AWS Harness Undercuts Claude Code by 77%

🔥 What happened AWS dropped Strands harness: a general-purpose agent harness, Apache 2.0, Python and TypeScript, live in one line of code. On Terminal-Bench 2.1 it ran $56 per task; Claude Code burned $248. 💡 Why it matters Same model (Fable 5), yet Strands hit 69.7% accuracy vs Claude Code's 61.8% — 77% cheaper and still ahead. The edge isn't magic: 1,500-token truncation, compaction at 85% context, in-loop recovery. ⚡ Our take The real story: the harness, not the model, decides your bill. Anyone still torching tokens in 2026 hasn't fixed their loop.