Claude Opus 5 Matches Fable 5 at Half the Price, With an Effort Dial. Agent Economics Just Changed

Anthropic launched Claude Opus 5 at $5 per million input tokens and $25 per million output, half the price of Fable 5, which it nearly matches on frontier benchmarks. It also shipped a per-turn effort control, a dial that sets how hard the model thinks on each call. For a chatbot, that's a nice-to-have. For an agent that makes dozens of model calls to finish one task, a halved price plus controllable effort is a direct cut to cost-per-completed-task, which is the number that actually decides whether automating a workflow pays. The headline is the benchmark. The shift underneath is that you can now tune cost and reasoning per step.

Most coverage is scoring Opus 5 against Fable 5 and the rest of the frontier. The more useful question for anyone running agents in production is what a cheaper model with a reasoning dial does to the unit economics of the work.

Why an effort dial matters more for agents than for chat

A person in a chat window sends one prompt and reads one answer. Effort per call barely registers. An agent is different. Finishing a real back-office task, screening a candidate, reconciling an invoice, classifying a case, takes a chain of model calls, and they are not all equally hard.

Most steps are mechanical: parse a field, reformat a value, check a box. A few are where the reasoning actually decides the outcome: the judgment call, the edge case, the ambiguous match. Until now you paid roughly the same reasoning cost on every step. A per-turn effort control lets you spend deep reasoning only where it changes the answer, and run everything else cheap. You're right-sizing compute to the step instead of paying frontier effort to reformat a phone number.

Half the price changes what's worth automating

Cost-per-task is simple: number of calls times price per call. Agents run a lot of calls, so the price per call compounds fast. Halving it, and then dialing effort down on the cheap majority of steps, doesn't trim the bill a little. It resets which workflows are worth automating at all.

A workflow that was marginal at $10/$50, where the compute cost ate the savings, can clear the bar at $5/$25 with effort tuned per step. The economics, not the intelligence, are what just moved.

The lesson isn't "switch everything to Opus 5"

The reflex with any cheaper-and-nearly-as-good model is to route everything to it. That's still the wrong read, and the effort dial makes the point sharper. The job isn't to pick one model and one setting. It's to route each step to the right model at the right effort: cheap model, low effort for the mechanical majority; stronger model or higher effort for the few steps that decide the outcome. Opus 5 adds a second knob to a decision that was already about matching cost to the task.

That's exactly what a model-agnostic platform is for. When the price-performance leader shifts again next quarter, and it will, you want to change a setting, not re-plumb the workflow.

What to actually do about it

• Measure cost-per-completed-task, not price-per-token. The token price is a headline; the number that matters is what one finished unit of work costs end to end.

• Dial effort down on the mechanical steps. Most of an agent's calls don't need frontier reasoning. Spend it where the judgment lives.

• Keep the expensive effort for the decisions. The edge cases and ambiguous calls are where high effort pays for itself.

• Stay model-agnostic. Today it's Opus 5 at half the price. The point isn't this model, it's the ability to route cost and reasoning per step without rebuilding.

The story isn't that Anthropic shipped a cheaper flagship. It's that agents made cost-per-task the metric that matters, and Opus 5 just handed you two levers to lower it. The teams that win the economics won't be the ones on the cheapest model. They'll be the ones routing each step to the right cost.