AGENT DESIGN · COST CONTROL

Effort dial

Thinking Machines Lab shipped its first model this week, Inkling, with a dial for controllable thinking effort: the same model trading reasoning depth for latency and cost instead of switching to a cheaper tier. This is a small working demo of that shape, one coding task run at four effort levels, no real LLM behind it, just the pattern.

Task

Validate a Philippine mobile number typed into an intake form, and format it to E.164 for the SMS provider.

Thinking effort

Output at medium effort

function toE164(input) {
  const digits = input.replace(/\D/g, "");
  const local = digits.startsWith("63")
    ? digits.slice(2)
    : digits.replace(/^0/, "");

  if (!/^9\d{9}$/.test(local)) {
    throw new InvalidMobileNumberError(input);
  }
  return `+63${local}`;
}

· Handles both 0-prefixed and 63-prefixed input.
· Enforces the PH mobile 9xxxxxxxxx shape, throws a typed error instead of returning null.

SHIP TO A CLIENT? Yes, this is what I'd want in most client intake forms.

Cost of this effort level

TOKENS

430 tok

LATENCY

1.6s

COST

$0.0018

Same model, same prompt, only the effort dial moved. No model switch, no separate cheap-tier integration to maintain.

WHY THIS MATTERS FOR CLIENT WORK

Most agent calls in a client build do not need frontier reasoning burned on every request: a copy tweak, a one-off script, a straightforward CRUD route. The failure mode is either always paying for max effort out of habit, or always running cheap and eating the edge-case bugs that shows up later. A dial that stays on the same model, instead of a manual switch between model tiers, makes it cheap to match effort to what the task actually needs, and to turn it up only for the calls that touch money, production data, or something a client will notice.

Built 16 July 2026 · pattern sourced from Thinking Machines' Inkling launch

THINKING MACHINES →BACK HOME →