Everyone will quote the leaderboard line from the Opus 5 announcement — near the frontier of the top-tier model at half the price. The number that actually changes how I’d run this in production is the effort setting: a dial that trades intelligence for token conservation on the same model, per request. That’s not a marketing knob, it’s the missing primitive for latency budgets and cost SLAs in an agent loop.

Most agent frameworks fake this today by routing between different models — a cheap one for tool-call plumbing, an expensive one for the hard reasoning hop. A single model with an explicit effort axis collapses that routing logic into one call and makes the cost curve legible instead of emergent.

• 🎯 Effort is per-call, so you can spend on the planning step and starve the 40 boilerplate tool calls around it
• 📊 Frontier-Bench v0.1: more than doubles the prior Opus at a lower cost per task — the cost-down, not just capability-up, is the story
• ⚡ CursorBench 3.2 at max effort lands within 0.5% of the top model at half the cost per task
• 🔍 OSWorld 2.0 computer-use wins at any given cost point — relevant if your agents drive real UIs
• ⚠️ Honest boundary: it’s explicitly behind another model on cybersecurity tasks — a rare bit of capability-gap disclosure in a launch post

The HN thread is already litigating benchmark validity, which misses the operational point. A frontier score you can’t afford at your QPS is a demo; a dial that lets you pick your point on the intelligence/cost curve per request is an ops feature.

If effort settings become standard across providers, does model routing as a discipline mostly evaporate — or just move up a layer into “how much effort does this sub-task deserve”?