✅ Now That Sonnet 5 Beats Opus on Some Benchmarks, How Do You Actually Choose?

Sonnet 5's Terminal-Bench win over Opus 4.8 complicates what used to be a simpler rule of thumb ("Opus for hard problems, Sonnet for everything else"). A more useful framing post-launch: pick based on task shape, not a blanket tier preference.

• Long, autonomous, tool-heavy runs (terminal orchestration, multi-step agentic workflows) — Sonnet 5 is now a legitimate first choice, not just a cost-saving fallback.
• Deep single-shot reasoning (a hard proof, an ambiguous architecture decision with no clear iterative feedback loop) — Opus 4.8 still generally has the edge where the benchmarks show it does.
• When in doubt, benchmark your own workload — published benchmarks are directional, not a substitute for testing against your actual task distribution, especially now that the gap between tiers has narrowed enough that general advice matters less.

✅ A Simple Pattern for Routing Requests to the Cheapest Model That Works

With the pricing gap between Sonnet and Opus tiers now wider than the capability gap on many tasks, a routing layer that classifies request difficulty before picking a model pays for itself quickly at scale. A minimal version: run a fast, cheap classification pass (even Haiku-tier) to estimate task complexity, then route accordingly.

def route_model(task_complexity_score):
    if task_complexity_score < 0.3:
        return "claude-haiku-4-5"
    elif task_complexity_score < 0.7:
        return "claude-sonnet-5"
    else:
        return "claude-opus-4-8"