The headline in the Sonnet 5 announcement — performance close to the flagship at a lower price — is a routing decision disguised as a model release. For anyone running an agent loop in production, the model you default to is a cost curve, not a leaderboard rank.

• 🎯 The default-model choice is the budget. An agent that takes 40 tool-call turns pays the per-token gap 40 times over. A mid-tier model that lands within a few points of the flagship changes which steps you’re willing to run autonomously.
• 📊 Benchmarks don’t tell you the tail. Aggregate agentic scores hide where a cheaper model loops, retries, or bails early. That only shows up in your own eval harness, on your own traces.
• ⚡ Escalation beats picking one model. The pattern that survives contact with production is cheap-model-first with escalation to a flagship on low confidence or repeated failure — a narrowed gap just means the first tier does more before it escalates.
• ⚠️ “Lower cyber capability” reads as a deployment feature. The system card notes reduced cybersecurity ability; for enterprise agents holding shell access, a narrower blast radius is a selling point, not a regression.
• 🔍 Re-baseline before you migrate. A model that’s cheaper per token but changes tool-call behavior can cost more end-to-end. Measure turns-to-completion, not price-per-million.

The HN discussion is mostly benchmark-versus-benchmark. The question I’d answer before switching anything: does Sonnet 5 finish my hardest agentic traces in fewer turns, or just cheaper per turn that I then burn back on retries?