Your AI Just Got an Effort Dial. Here's When to Turn It Up.

Most of the time, your AI doesn't need to think hard. A reworded subject line, a first-draft reply to a routine customer query, a quick formatting pass — fast and cheap is fine. A few tasks are different: diagnosing a process problem, pressure-testing a strategy, checking whether a plan has a fatal flaw. Those benefit from more careful reasoning. Claude Opus 4.8, released May 28, is the first mainstream AI tool that lets non-technical users choose which kind of output they're getting — explicitly, before the model starts.

What shipped

Anthropic released effort controls with Claude Opus 4.8 on May 28. Fast mode runs 2.5x faster and costs 3x less. Deep mode applies more thorough reasoning — Anthropic reports the model is 4x less likely to let input flaws pass without flagging them. The Neuron covered the rollout on May 29, noting the controls come alongside dynamic multi-agent workflows. For professionals who don't build AI systems for a living, the multi-agent story is secondary. The primary change is simpler: you now tell the model how hard to work before it starts.

What changes on Monday

Two things to do differently.

First, stop treating all tasks as equally worthy of thorough AI reasoning. Most of what you send — rephrasing, drafting, formatting, summarising — doesn't require deep analysis. Running everything at maximum effort costs more time and money without improving results for tasks that don't need it.

Second, when a task actually does require reasoning — diagnosing a problem, evaluating options, finding what you're missing in a plan — use deep mode and frame the task to match. A vague prompt in deep mode is still a vague prompt. The model will reason carefully about the wrong problem if you don't give it the right one.

Sophie: the brand strategy problem

Sophie is a brand manager at a 35-person SaaS startup. She uses AI daily: social copy, customer review responses, internal comms, quarterly brand strategy briefs.

Before effort controls, everything went through the same pipeline at the same depth. A subject line test and a strategic positioning brief got identical reasoning treatment — which meant either the subject line was over-engineered or the strategy was under-reasoned.

With effort controls, she's made a simple split: creative and copy tasks go fast; anything involving her own assumptions goes deep. Her most useful application: a competitor positioning brief where she asked the model to find the weakest part of her draft strategy before she brought it to her co-founder. In fast mode, that prompt produces minor copy suggestions. In deep mode, it surfaced a segment overlap she hadn't considered — the kind of catch that changes the brief, not just the wording.

Marcus: the procurement diagnosis problem

Marcus is an operations manager at a 200-person industrial equipment manufacturer. He's been trying to diagnose why one procurement category — specialist fasteners — keeps missing its 14-day fulfilment target while every other category hits it.

He'd asked AI about this before and gotten a standard list: supplier lead time, demand forecasting, safety stock. Nothing specific to his situation.

This time he used deep mode, gave the model the actual data — fulfilment times by category, supplier count, demand volatility — and asked it to reason through which combination of factors most likely explains the gap. The model identified that the fastener category had the lowest supplier count and the highest demand variance of any underperforming category across comparable facilities. That wasn't in the generic list. It required working through the problem rather than pattern-matching to a standard answer.

The tradeoff is now visible

The effort dial doesn't make AI smarter. It makes an existing tradeoff explicit — fast and cheap, or slow and thorough — so you can match the mode to what the task actually needs. Most tasks don't warrant deep reasoning. A few do. Knowing the difference before you hit send is the skill.