Every AI Tool Now Asks How Hard to Think. Most People Don't Answer.

Every AI tool you use shipped the same new control this year, worded three different ways: a setting that decides how hard the model thinks before it answers. Claude calls it effort. Gemini split into three tiers and dropped the old technical knobs entirely. The setting exists in every major product now — and almost nobody has a rule for when to touch it.

Four announcements, one design decision

It started in late May, when Claude Opus 4.8 shipped explicit effort controls: fast mode at a third of the cost, deep mode markedly more likely to catch a flaw in your input before it becomes your problem. By late July, Claude Opus 5 turned that into a three-position toggle, and Google deprecated temperature, top-p, and top-k across Gemini rather than keep offering the old dials alongside the new tiers. A few weeks before that, GPT-5.5 Instant showed what happens when a model gets better at doing exactly what you asked and nothing more — shorter, more accurate answers that stopped padding around a vague prompt. By August, DeepSeek and OpenAI had cut prices low enough that a capable model runs a fraction of a cent per million tokens, which turned "which tier" from a cost question into a judgment one.

Four companies, four separate announcements, the same underlying decision: the provider no longer guesses how much reasoning your task needs. You tell it.

The rule that answers it

Picking the wrong tier is now something you do, not something that happens to you — and "whichever setting worked last time" isn't a rule. This one is:

Routine tasks — reformatting, first-pass drafts, rephrasing a subject line — get the fastest, cheapest tier. Nothing here benefits from deeper reasoning; you'd be paying for elaboration you delete anyway.

Analytical tasks — a summary feeding a real decision, a first-round competitive read — get the default or mid tier. Enough reasoning to be useful, not enough to hedge a first draft into mush.

High-stakes tasks — anything you'd stake your name on, anything where a wrong number is expensive to unwind — get the top tier, every time, cost aside.

None of this replaces framing the task well. A deep-mode prompt typed in thirty seconds between meetings is still a vague prompt — tier selection is the second decision, not a substitute for the first.

A logistics ops lead learns the difference

Tom runs operations for a 90-person freight brokerage. His team's most common AI task is a weekly carrier scorecard — pull last week's on-time numbers, format them for the ops review. He'd been running that, and everything else, through the same mid-tier default, because that's the setting the team standardized on when they rolled the tool out.

The scorecard was never the problem. The quarterly carrier-risk memo — the one that recommends which vendors to renew — had been running through that same default. Switched to the top tier, the memo flagged a carrier whose on-time rate looked fine in aggregate but had cratered on the two lanes carrying 40% of his volume, a pattern the mid-tier version had summarized past.

A grants manager learns the other direction

Angela manages grants for a 25-person nonprofit. After reading about effort controls, she did the opposite of Tom: ran everything on the highest tier, reasoning that more careful was always safer. Her weekly batch of funder acknowledgment letters — thirty near-identical thank-yous with a name and amount swapped in — took longer to generate and cost more, for output that read no better than the fast-tier version.

She now reserves the top tier for the one task that earns it: the annual compliance narrative that goes to funders directly, where a misstated grant term is the kind of error that costs the relationship. Everything else runs fast. The letters didn't get worse. The budget she freed up covers running the compliance narrative on the expensive tier twice — a draft, then an adversarial second pass — for less than she'd been spending running everything on one setting.

The one thing

The provider decides how many tiers exist. You decide which one a task has earned. That's a five-second decision when you make it on purpose, and an expensive default when you don't.