Opus 4.8 and the Quiet Win of Fewer Tool Calls

The headline on Claude Opus 4.8 is the benchmark bump — 84% on Online-Mind2Web, first model to break 10% on a legal agent all-pass standard, gains on CursorBench. Those are the numbers everyone screenshots. The line I actually care about for production is buried lower: it uses fewer steps for the same intelligence on tool calling.

If you operate agents, you know step count is latency and cost. Every extra tool call is another round trip, another chance to fall out of the cache, another place for a retry to fire. A model that reaches the same answer in fewer hops moves your P99 and your bill more than a couple of points on a leaderboard ever will. Same story with the Messages API now accepting system entries mid-task without breaking the prompt cache — unglamorous, but that’s a real money lever when you’re caching aggressively.

• ⚡ Fewer tool-call steps → shorter agent traces, tighter latency budgets.
• 🎯 ~4x less likely to let code flaws pass — useful, but I’d want to see it against my eval set, not the vendor’s.
• 💡 Effort slider — explicit depth/speed tradeoff is the right primitive; it should have been a first-class API knob a year ago.
• 📊 Same price as 4.7 ($5/$25 per M tokens) — capability up, cost flat is the trend that actually compounds.

I’ll take vendor benchmarks as a hypothesis, not a result. The agentic numbers only matter once they survive contact with your own eval harness and your own tools — that’s where models that look great in demos quietly regress.

The HN discussion is mostly “is this a real jump or a point release?” Honestly, for agent workloads the boring efficiency wins beat the benchmark deltas. When you upgrade a model, do you re-run your eval suite — or just trust the release notes?