There’s a second dial on your AI model, and most people never touch it. They pick a model, and then quietly leave its reasoning effort pinned to maximum — because more thinking sounds like more quality, and who wants less quality?

Here’s the uncomfortable part: on a huge share of the work you do, that extra thinking changes nothing about the answer. It just costs more to produce the same output. You’re paying the model to deliberate over a decision that was never in doubt.

Model choice — the subject of the first article — is the big lever. Effort level is the quiet one sitting right next to it, and once you see it you can’t unsee how much of your budget it’s been eating.

The evergreen idea: effort is a separate lever from model

Every current frontier model ships with a way to control how hard it thinks before it answers: Claude’s extended-thinking budget and effort setting, GPT-5.6’s reasoning-effort control, the “thinking” toggles on the challenger reasoning models. This is independent of which model you picked. You can run a cheap model at high effort or an expensive one at low effort — they’re two separate knobs.

And effort has a shape you need to internalise: it pays off on hard problems and does nothing on easy ones, but it costs more on both.

  • On a genuinely hard task — a subtle concurrency bug, an architectural trade-off, a proof that has to be right — more reasoning genuinely finds better answers. The value keeps climbing as you turn the dial up. Spend it.
  • On an easy or determined task — rename this, format that, write the obvious CRUD handler — the answer is already fixed the moment you ask. Extra reasoning produces the same result after burning three times the thinking tokens. The value line went flat; only the cost line kept climbing.

The mistake isn’t using high effort. It’s using high effort by default, which means paying the hard-task price on the mountain of easy tasks that make up most of a real project.

Reasoning effort versus value: hard tasks keep paying off, easy tasks stop early while cost keeps climbing Value Low Medium High Max Reasoning effort → Easy task: flat past Medium — you buy cost, not value Cost / budget drawn down (always climbs with effort): Hard task Easy task

The current state (as of 19 July 2026)

The controls differ by vendor, but they all do the same job. Details move, so verify against current docs before you script anything around them.

As of 2026-07-19:

  • Claude (Fable 5, Opus 4.8, Sonnet 5): extended thinking can be toggled and given a token budget; you cap how many tokens the model may spend reasoning before it answers. More budget, more deliberation, more cost.
  • OpenAI GPT-5.6 (Sol / Terra / Luna): a reasoning-effort setting (and a separate speed setting). Higher effort spends more reasoning tokens and, on message-metered plans, draws your allowance down faster.
  • Challenger reasoning models (DeepSeek’s reasoning line, GLM’s thinking modes, others) expose the same idea under different names.

How effort hits each meter (see the metering article): - Token meter: reasoning tokens are billable output tokens. High effort literally multiplies the tokens you pay for on every call. - Message meter: high-reasoning requests count harder against your cap — community reports consistently describe the weekly pool draining “exponentially faster” when effort is maxed. - Compute-hour meter: more thinking is more throughput, so it eats the pool the same way long context does.

On every meter, effort is a cost multiplier. The only question is whether the task pays it back in quality.

Two tasks, one dial

Take a mechanical one first: “rename getUser to fetchUser across the module and update the imports.” At low effort the model does it correctly in seconds. At max effort it does the identical edit, having spent a few thousand reasoning tokens convincing itself of something that was never ambiguous. Same diff, several times the cost. Run that pattern across a hundred small edits in a project and the waste is not a rounding error — it’s a meaningful slice of your monthly budget, spent on deliberation nobody needed.

Now a hard one: “this endpoint intermittently returns stale data under load — find out why.” Here low effort will cheerfully hand you a plausible, wrong answer, and you’ll waste an hour chasing it. High effort is where the model actually traces the race condition, weighs the caching layer against the transaction boundary, and gets it right. The reasoning tokens bought you a correct answer and saved the hour. That’s effort earning its cost.

Same dial, opposite verdict. The skill is telling the two situations apart before you spend — and defaulting to medium so that the expensive setting is a deliberate choice, not an accident of configuration.

Where the challengers fit

The cheap reasoning models make effort discipline even more valuable, not less. A model like DeepSeek’s reasoning line is inexpensive per token, but crank its thinking wide open on easy work and you throw away the very cost advantage you chose it for. The right move is the same everywhere: cheap model and low effort for determined work; escalate model and effort together only when the problem earns it. Effort discipline is what keeps a cheap model actually cheap.

The playbook

  • Default to medium, not max. Make high effort a decision you make, not a setting you inherit.
  • Turn thinking off for mechanical work. Renames, formatting, boilerplate, applying a clear fix list — these need production, not deliberation.
  • Cap the thinking budget where you can. On Claude, set a sane extended-thinking token ceiling rather than an open-ended one. You rarely need the model to think to its limit.
  • Escalate model and effort together, deliberately. The genuinely hard 20% deserves both. The other 80% deserves neither.
  • On message-metered plans, watch effort especially. It’s the fastest way to empty a weekly allowance without noticing.
  • Re-check the controls. Effort settings and their billing get renamed and retuned; confirm current behaviour before automating around it.

The Khula Take

The industry’s reflex is “more reasoning is better” — vendors market bigger thinking budgets like horsepower, and builders leave the dial on max to feel safe. We think that’s backwards. High effort is not a quality setting you should default to; it’s a cost you should be trying to avoid needing. The goal isn’t a model that deliberates beautifully over every request — it’s a workflow where most requests are so well-scoped they don’t require deliberation at all.

For a South African business, that reframing is where the money is. Instead of paying a premium model to think hard about a vague instruction, you invest a little upfront in making the instruction unambiguous — clear specs, deterministic scaffolding, tight prompts — so the AI can answer cheaply and correctly. Applied AI that grows a business isn’t the one that reasons the hardest; it’s the one engineered so the hard reasoning is rarely necessary. Every turn you can safely run at low effort is margin you keep — and margin, reinvested, is how a small operation out-ships a bigger one.

Where this is heading

Choosing an effort level per task, on every plan, is one more judgement call to make hundreds of times a week. It’s precisely the kind of decision the Token Economy harness is designed to make for you — reading the task, setting model and effort to match, and never leaving the dial on max by accident. The estimator, in turn, will price a build with effort as an explicit variable, because on a message meter it’s often the difference between “fits the week” and “throttled by Wednesday.”

Until then, the discipline costs nothing and starts today: turn the dial down to medium, and only turn it up when the task looks you in the eye and asks for it. Confirm the current effort controls against your vendor’s docs before you build anything around them — like everything in this series, the specifics have a short shelf life.

Sources (July 2026): - Anthropic extended thinking & pricing — platform.claude.com/docs - OpenAI GPT-5.6 reasoning-effort & tiers — OpenAI Help Center - Effort/limit drain (community/aggregator) — eesel: GPT-5.6 pricing

Effort controls and their billing change frequently. Confirm current behaviour against the vendor docs before relying on any figure here.