Anthropic dropped Claude Opus 5 today. The headline: near-Fable-5 intelligence at half the cost. The subtext: the unit economics of agentic work just rewired.
Opus 5 isn't a frontier model — that's still Fable 5's lane. But it closes the gap to within 0.5% of Fable 5's peak score on Frontier-Bench v0.1 at max effort, at half the cost per task. It more than doubles Opus 4.8's performance on the same benchmark, at a lower cost per task. That's not a linear improvement. That's a regime change in what's economically viable to run.
Where the numbers land
Coding: On Frontier-Bench, Opus 5 surpasses every model at any cost tier. On Zapier AutomationBench, its pass rate is ~1.5× the next-best model at the same cost per task.
Computer use: Opus 5 outperforms every other model at any given cost, beating Fable 5's best result at just over a third of the cost.
Scientific research: Better than Opus 4.8 on every life sciences evaluation — structural biology, genomics, statistical reasoning.
Self-correction: On one Frontier-Bench task, Opus 5 was given a drawing of a machine part and asked to write code to rebuild it in FreeCAD — intentionally with no way to read the drawing. It wrote its own computer vision pipeline to pull geometry from raw pixels, then reconstructed the full part.
What this means
Opus 4.8 was already the workhorse of serious agent deployments. Opus 5 at the same price point means any workflow that was cost-sensitive at 4.8 — iterative debugging, long-horizon research, multi-step planning — suddenly gets 2× the performance budget. For agents that loop over tool calls, the ratio of success per dollar is the metric that matters, and Opus 5 shifts that curve sharply.
Testing the limits behavior — building its own CV pipeline when blocked, finding a real bug that a community patch missed — hints at something deeper. Opus 5 doesn't just follow instructions better. It pushes back when the approach is wrong. Multiple early-access testers reported the model arguing against flawed designs, not blindly executing.
The Opus tier has always been the practical choice — capable without the frontier premium. Opus 5 makes that calculus lopsided. If you're paying for agent inference, recalculate.