Leaving model effort at maximum for every task means avoiding a decision I still own. Anthropic added effort controls in Claude with the Opus 4.8 announcement on May 28, 2026. I like having the control. Dragging a selector all the way to the right, though, tells me very little about whether I have assigned the right amount of work.
In September 2026, during a conversation about execution, I asked for model selection to follow the task. The request was straightforward: use what makes sense for the job. Looking back at the May announcement, I apply that same expectation to effort. Choosing a model and deciding how much it should investigate both need a reason.
A wording change with a settled request deserves an edit and a check. A change across several parts of a codebase can need investigation first. Treating both alike looks like an expensive way to avoid thinking. The cost includes my time waiting while an agent develops another plan for something that was already decided.
Extra effort needs to buy something I can identify. Finding a hidden dependency is worth waiting for. Restating the uncertainty in four paragraphs does little for delivery. The work returned to me also counts: if I have to explain the request again or reconstruct what happened, that quick response was less cheap than it looked.
For a comparison, I want a bounded task and a definition of completion written before the answer arrives. Then I can examine what changed with the setting. Did review become easier? Did the agent catch something that would otherwise have slipped through? Did it resolve the uncertainty that justified investigating? Answering those questions takes work, but it produces a decision I can defend.
I also want the setting to follow changes in the task. A small edit can expose a serious constraint. An investigation can finish sooner than expected. Thinking after the uncertainty is resolved just delays execution. A brief update about the remaining question is enough; I do not need a transcript of the reasoning dumped into the conversation.
For the next comparison, I want a concrete benefit before keeping high effort as the default. I will separate clear tasks from investigative ones and judge the delivery alongside the review left for me. If the only explanation is that the button goes up to maximum, nobody has actually made the decision yet.