Opus 5.5 and the Effort Dial: What One App at Every Level Actually Measures

01 The Experiment: One Spec, Every Effort Level

Claude's newest frontier release ships with a dial most users never think about: the effort setting, which controls how hard the model works before answering. Nate Herk's experiment is the cleanest public demonstration of what that dial does: he gave Opus 5.5 the same application spec at every level and watched five different builds emerge from one prompt. Same model, same request, same tools ' different budgets of thought.

The design matters because it isolates one variable. Benchmarks compare models; this compares a single model against itself across its own effort settings, which is increasingly how real teams configure agents in production. The spread between the builds is therefore not noise ' it is the price of attention, made visible.

02 What the Dial Actually Changes

Effort is best understood as a budget for metacognition. At low settings the model reads the spec, drafts, and ships. At high settings it plans an architecture, writes to-do lists, builds component by component, tests its own output, and iterates when tests fail ' all of which consumes reasoning tokens, tool calls, and wall-clock time. The visible work product changes less than the invisible process underneath it.

This is why effort levels feel different in agentic workflows than in chat. A conversational answer at low effort is usually fine; a multi-file application build at low effort skips the self-correction loop where most quality comes from. The dial does not make the model smarter ' it decides how many chances it gets to catch itself.

03 Quality Versus Tokens: Reading the Spread Honestly

The outcome spread, idealized in Chart 2, follows a pattern familiar from every scaling result: quality rises gently and cost rises steeply. The max-effort build completed more of the spec, handled edge cases the others missed, and needed fewer human corrections afterward. But the medium-effort build captured most of the value at a fraction of the tokens, and even the low-effort build produced a working skeleton.

The honest reading is not that high effort is wasteful ' it is that the marginal token buys less than the average one. The gap between max and high is small in quality and enormous in cost; the gap between minimal and low is large in quality and trivial in cost. Rational usage lives at the knee of that curve, and most users have no idea where the knee is.

04 Latency and Price: The Tradeoffs Users Feel

Cost is only half the bill. A max-effort build takes proportionally longer ' the thinking loop cannot be parallelized away ' which turns the dial into a latency control as much as a quality control. For interactive work the extra minutes are noticeable; for batch agent runs they are invisible. This is why the same dial serves opposite use cases: chat wants low effort for responsiveness, overnight pipelines want max effort because nobody is watching.

Pricing ties the two together. Token-based billing means the effort setting is literally a price setting, and the published math is unforgiving: the max-effort build in this experiment plausibly cost an order of magnitude more than the minimal one. Subscription plans hide this by capping usage; API bills do not. The dial is where that abstraction ends.

05 Where Low Effort Wins and Where It Breaks

Low effort has a legitimate kingdom: well-specified tasks with clear success criteria, small diffs, routine refactors, formatting, translation, classification. When the space of correct answers is narrow, thinking longer mostly finds the same answer more expensively. Teams that mandate high effort everywhere are burning budget to delay obvious responses.

The breaks appear where specification is incomplete. Low-effort builds in this experiment made early architectural commitments that later sections of the app had to route around ' the model answered the question asked instead of the system implied. That is the signature failure of under-thinking: locally fluent, globally brittle. The fix is not always more effort; often it is a better spec. But when the spec cannot be improved, effort is what remains.

06 The Routing Future: Attention Billed by the Task

The experiment previews where pricing is heading. Frontier labs are converging on usage-based billing with model routing: cheap models for easy calls, frontier models at high effort for hard ones, decided automatically. The effort dial is the manual version of that router, and its spread is the spread the routers will monetize. Attention is becoming a metered utility, billed like electricity ' by the task, at rates that vary with how hard the thinking has to be.

For buyers, the skill of the next few years is matching effort to stakes. The teams that treat every request as a max-effort request will pay supercomputer prices for secretary work; the teams that default everything to low will ship brittle software. The dial makes the tradeoff explicit. That is not a billing nuisance ' it is the first honest price tag ever attached to machine reasoning.

07 How to Read Effort Demos Without Being Sold To

A caution: effort-level demonstrations are marketing-adjacent by construction. The presenter chose the task, the spec, and the camera angles, and a same-app-at-every-level format flatters the vendor twice ' once by showing the top setting succeeding, and again by implying the dial is the product's best feature. Watch for what is held fixed: the spec quality, the evaluation criteria, the number of retries each build received.

The durable takeaway survives the caveats. Effort settings are real, their cost curve is steep, and the quality curve is gentle ' which means the highest-value skill in AI tooling is no longer prompting but budgeting. The models have learned to think longer. The users now have to learn when that is worth paying for.