Anthropic Explained When to Tinker with Effort Level in Claude Code (Spoiler: Not Often)

In the world of AI coding assistants, it’s easy to treat every dial like a magic button that makes outputs instantly better. Claude Code’s two main controls — model selection and effort level — look similar at first glance. But Anthropic’s own team has now published a clear breakdown showing they do fundamentally different things. And their core advice? Leave the knobs alone most of the time.

Model = Frozen Knowledge, Not Magic

Anthropic Explained When to Tinker with Effort Level in Claude Code (Spoiler: Not Often) A model is simply a fixed set of weights trained once. Everything Claude “knows” about TypeScript patterns, Go idioms, popular frameworks, or edge-case bugs lives in those weights. Your prompt, `CLAUDE.md` file, conversation history, and open files are all just context — they steer the model for that one request but don’t teach it anything new.

Changing the model swaps out which set of frozen weights handles your request. Larger models (Fable → Opus → Sonnet) have stronger baseline capabilities on benchmarks and can handle more ambiguity or subtle problems. But they also cost more per token. There’s no point paying for frontier intelligence on routine refactors.

Effort = How Much Work It’s Willing to Do

Anthropic Explained When to Tinker with Effort Level in Claude Code (Spoiler: Not Often) Effort level is not simply “think harder” or “take more time.” It’s a behavioral signal baked into training that controls *how much work* Claude performs before checking back in with you.

This includes:

How many files it reads;
Whether it runs tests or double-checks;
How far it pushes through multi-step tasks;
How thorough its plan is and how many verification steps it takes.

Crucially, effort is passed as an additional input token alongside your prompt. The model was explicitly trained on different effort behaviors, so it’s not an after-the-fact hack — it’s part of the model’s learned policy.

At higher effort, the same task can consume roughly 7× more tokens to reach higher confidence. But it’s a ceiling, not a quota. If the first debugging hypothesis nails the bug, Claude skips the rest and tells you why — Anthropic actively trains against wasteful overthinking.

The Practical Diagnosis Heuristic

When something goes wrong, don’t immediately reach for the sliders. First, fix your context (vague prompt, missing `CLAUDE.md`, wrong tools).

Then ask the key question:

Did Claude read everything, try honestly, and still fail? → Capability issue. Switch to a bigger model.
Did it skip files, bail on tests, or abandon a refactor halfway? → Effort issue. Raise the effort level.

Anthropic’s own guidance: default effort levels are calibrated for what most users would reasonably want to spend on a typical task. For 80% of work, you don’t need to touch them.

Helpful Analogy: Specialists, Experts, and Generalists

Think of the models this way:

Fable — the narrow specialist who has seen obscure problems almost no one else has. Even on low effort, it spots things others miss.
Opus — the deep expert.
Sonnet — the strong, reliable generalist.

Then layer on effort as time spent:

Opus on low effort ≈ five minutes with a world-class expert (quick, insightful read but not exhaustive).
Sonnet on high effort ≈ giving a talented senior engineer the whole day to read every file, run tests, and understand your codebase deeply.

Neither is universally superior. Real tasks need the right mix of raw capability and thoroughness.

The Economics: Bigger Isn’t Always More Expensive

Counter-intuitively, on complex multi-step tasks a larger model can be cheaper per successful outcome. A smaller model may thrash through many short, failing iterations until it hits its capability ceiling. A bigger model reaches a working solution in fewer turns. On routine work, though, the bigger model is just burning money for no gain.

Bottom Line from Anthropic Itself

“Don’t pay us for a model the task doesn’t need.”

Start with defaults. Diagnose failures honestly: didn’t know versus didn’t try. Tune effort as a broad preference for your workflow type rather than per-task micromanagement. And remember — the best lever is usually better context and clearer instructions, not another twist of the dials.

Claude Code’s power comes from understanding these two axes clearly. Model sets the height of the ceiling. Effort sets how aggressively you climb toward it. Most of the time, the defaults Anthropic ships are already pretty close to optimal.

---