Every model launch this month led with a leaderboard. Inkling — the first open-weights model from Thinking Machines Lab, a 975-billion-parameter mixture-of-experts (41B active per token) released mid-July under Apache 2.0 — has the leaderboard too: Thinking Machines reports 77.6% on SWE-bench Verified at maximum effort. But the number that should change how a founder budgets isn't on the chart. It's a single request-time knob, reported to run from 0.2 to 0.99, that sets how many reasoning tokens the model spends before it answers.

If you only remember one thing: on a reasoning model, the reasoning tokens are the bill — and Inkling makes that spend a number you set, not a surprise you discover.

The leaderboard is measured at a setting you'd never ship#

Here's the sleight of hand in every reasoning-model announcement. That 77.6% is posted at maximum effort — the model thinking as long as it possibly can before answering. That's the right way to win a benchmark and the wrong way to run production traffic. Max effort means max output tokens, and output tokens are where the invoice actually comes from. A leaderboard tells you what a model can do when cost is no object. Your unit economics are decided at whatever effort you can actually afford on every call.

Most models hide that gap. You get one behavior, priced one way, and the reasoning spend is emergent — you find out how much the model "thought" when the bill lands. Inkling exposes the gap as an input. That reframes the reasoning-effort-vs-thinking-budget question from a config footnote into the primary economic control on the model.

The dial is the product#

Think about what a per-call effort knob actually buys a small team. It's the same lever we already reach for everywhere else in the agent stack — the model-escalation ladder that sends cheap calls to cheap models and hard ones up-tier, context editing that evicts re-fetchable tokens so the window (and the bill) stays small. Those are all ways of not paying for compute you don't need. A 0.2-to-0.99 effort dial does the same thing inside a single model: it lets you buy exactly the amount of reasoning a given task is worth.

So the move isn't "should I switch to Inkling." It's: set an effort floor per task tier and stop overpaying.

Getting that assignment right is worth more than a few points of benchmark. A workload that runs everything at max effort because that's the default is lighting money on fire on its cheapest requests.

A benchmark score is measured once, in a lab, at max effort. Your bill is measured a million times, in production, at whatever effort you shipped. The dial is where those two numbers diverge — or don't.

Why Apache 2.0 makes the knob durable#

A closed frontier model can expose effort levels too. The difference is ownership. Rent the effort dial and you get the vendor's tiers at the vendor's price, with a default you can't see. Inkling's weights are Apache 2.0 — commercial use, fine-tuning, and redistribution, royalty-free — so you own the whole curve. You bake a house-default effort into your own serving layer, quantize to fit the GPUs you rent or own, and measure the token cost of each setting directly instead of reverse-engineering it from an invoice. That's the same reason the open-weight wave matters for cost at all: the control moves from the provider's pricing page to your config.

The token-efficiency claim is the one worth testing yourself. Thinking Machines pitches Inkling as matching rivals' scores using far fewer tokens — and tokens, not scores, are what you pay for. If that holds on your tasks, the effort dial isn't just a convenience; it's a structurally cheaper cost curve.

What to actually do#

Don't rip out your backend for a launch-week benchmark. Put Inkling behind the same abstraction as your other models and run your real workload through it at the effort levels you'd ship — not at max. Compare total serving cost, not the headline number. If you're already running open weights, the question is narrow and answerable: does Inkling's low-effort setting beat your current cheap tier on cost-per-acceptable-answer? That's a measurement, not a vibe — and it's the only comparison that shows up on the bill.

The industry spent this month arguing about who tops the chart. The more useful question Inkling raises is quieter: once reasoning is a dial, what effort does each of your calls actually deserve? Answer that per task tier and you'll cut spend without touching quality where it counts — which is the whole game when you're a team of one.