Anthropic recently took apart a dual-switch system most users never realized existed: Model (the brain) and Effort (the attitude) are two different things.
This piece argues a thesis: competition in AI coding is shifting from “whose model is stronger” to “who orchestrates agents better” — and those who master Model and Effort as independent levers will build a structural efficiency advantage.
An “AI Got Dumber” Scare Exposes the Ignored Switch
In March, a trust crisis erupted among Claude Code users. Developers found Claude Code suddenly “dumber” — skipping files it should read, not running tests, abandoning tasks halfway and asking for more information. GitHub filled with complaints. AMD’s AI lead Stella Laurenzo dug through logs of 6,852 sessions and measured Claude’s thinking volume down 67% from before February, concluding: “Claude can no longer be trusted with complex engineering work.”
Everyone first blamed their own prompts. Then came the realization: the problem was not theirs — on March 4, Anthropic had dropped Claude Code’s default Effort setting from high to medium. The cost of pressing latency had made the model “lazy.” After a month of blowback, Anthropic restored the default on April 7 and reset usage quotas for subscribers. But by then most users had only just learned that this switch was always at hand — quietly deciding how hard the AI would actually work for them. The deeper problem it exposed: the vast majority of users have exactly one control dimension for AI tools — swapping the model — when Anthropic’s architecture never had only one lever.
Model Versus Effort: Two Independent Levers
Model swaps the brain
Each model is a set of frozen weights. Its abilities and knowledge were welded shut the moment training ended: read-only, unchangeable. The prompts you feed at inference and the code you paste into context cannot alter those weights — you can steer, but not train. Swapping models means swapping an entire weight set for your job; it solves the problem of “can’t do it.” A library that did not exist at training time can be learned on the fly from pasted documentation, but only for that one request — the model remembers nothing and forgets instantly. And model inference works by token-level prediction: your code is split into tokens, each mapped to a vocabulary number — const is 1078, await is 2597 — and a 200-token response is 200 full computation passes. Your waiting time and your spend live mostly inside that loop.
Effort swaps the attitude
Many assume high Effort means “thinking a few seconds longer” — wrong. It governs how much work Claude invests in the task: how many files it reads, whether it runs tests, whether it performs extra verification, whether it pushes a multi-step job all the way through before returning. Low Effort Claude tends toward quick replies and asking you for more context, avoiding action where possible; high Effort Claude digs for information itself, makes extra tool calls, and finishes long task chains in one go. Anthropic’s official diagram: the same prompt at high Effort produces roughly 7× the tokens of low Effort — the surplus all spent on reading files, running verification, double-checking. Effort is not a speed dial. It is a behavioral signal of “to what standard is this job considered done.”
The counterintuitive core: small model + high Effort beats large model + low Effort
Here is the counterintuitive conclusion, and the analysis’s most important judgment: “it can’t” and “it didn’t try” are two entirely different failures requiring entirely different fixes. Anthropic’s official decision framework runs clean: Step one — when Claude fails, do not reach for the model. Check context first: is the prompt clear? Are the right tools available? Is CLAUDE.md configured? Most “AI got dumber” complaints root here. Step two — context is fine and it still fails. Then ask: can’t it, or didn’t it try? “Didn’t try” is easy to spot — skipped files, unrun tests, half-done refactors bouncing back as questions. That is Effort; raise it a notch. “Can’t” is different — context is complete, effort is visibly maximal, and it is still wrong; rephrasing fails again. No amount of Effort helps. That is Model; swap it.
Anthropic’s analogy: Sonnet is an all-rounder with a whole afternoon — reads your code end to end, runs everything twice; Opus is an expert giving you five minutes, bringing experience and instincts you lack; Fable is the specialist called when everyone else is stuck, spotting what others cannot. Hence the counterintuitive verdict: Sonnet at high Effort genuinely beats Opus at low Effort on many jobs.
The Dual-Lever Frame: From One-Dimensional Selection to Two-Dimensional Orchestration
| Model: low | Model: high | |
|---|---|---|
| Effort: low | Minimal investment, minimal output — simple queries and debugging | An expert unwilling to work — wasted resources |
| Effort: high | A small model at full effort — covers 80% of scenarios | The strongest combination — standard for hard problems |
The frame’s insight is not the quadrants themselves but the neglected fact they reveal: most AI-coding users still operate in one dimension — picking the model, ignoring the investment. Imagine a team with a prestige-school expert who slacks daily and a state-school graduate who hustles. In most projects — the ones not requiring Nobel-level insight — the graduate’s effort covers 80% of the need. Yet you pick the expert for his pedigree, then complain about output. That is how most people use Claude Code today. More notably, the latest Claude Code added multiple Effort tiers including a new “ultra” level — Anthropic itself has realized that spending more tokens on “verifying its own output” beats spending more parameters on “memorizing more knowledge” for certain scenarios.
What the Frame Explains Elsewhere
The dual-lever frame is not Claude Code-specific. It reveals a deeper application-layer trend: the model-capability ceiling is shifting from parameter count at training time to compute budget at inference time. Beyond coding — take an AI customer-service system: the same GPT-4-class model configured for “quick replies until human handoff” (low Effort) versus “understand the user’s deep intent and resolve autonomously” (high Effort) can differ more than swapping models would. In writing — a mid-tier model plus an instruction system of “think step by step, outline first, self-review after” often outproduces a stronger model’s unconstrained one-shot. The reason is not that the mid-tier model is smarter; it was given higher Effort. In enterprise agent deployment — 31% of enterprises ran agents in production as of 2026 — when managers complain agents are “unreliable,” how much is wrong-model versus insufficient effort configuration: iteration counts, verification steps, fallback strategies?
The dual-lever frame’s value: it converts the vague complaint “AI is unreliable” into two precise diagnostic questions — can’t it, or didn’t it try?
What To Do
Developers: stop. Next time Claude fails, adjust Effort before swapping models. On most daily tasks, Sonnet at high Effort is cheaper and more effective than jumping to Opus — reserve the model swap for genuinely hard problems that need its expertise. Team managers: reorder your AI tooling budget priorities: context quality above Effort configuration above model upgrades. The most common waste is buying the strongest model while ignoring the first two levers — reallocate the saved token budget to more iteration rounds rather than more parameters. AI product builders: do not ship a bare model picker. Offer an “investment level” or “rigor” control and teach users that the two dimensions are independent — that is both user education and product differentiation. Zooming out: if Anthropic’s official breakdown marks anything, it is that AI coding is leaving the “swap the brain” era for the “orchestration” era. Models will commoditize; the moat moves to who designs better agent workflows and who configures the right effort level at each task stage. That is not an algorithm problem — it is an engineering-management problem.
Information from Anthropic’s official blog, Synced reporting, and Claude Code documentation. The thesis and framework are independent analysis, not Anthropic’s official position.
