The effort dial

The reasoning control became a qualitative level, from low to max, but the old token budget did not disappear.

Part of the Getting good answers track on lAItest.

The setting used to be a number of tokens. Now it is a word: low, medium, high, xhigh, max.

And both still exist.

From budget to effort.

The first thinking controls were token budgets. You granted the model a number of thinking tokens and hoped you had guessed well: too low and it stopped mid-thought, too high and you paid for slack. The replacement is a qualitative level. You say how hard to work and the model decides what that costs in tokens. Several labs arrived at this same shape independently.

A common misconception

Commonly believed: Effort levels replaced token budgets. The old parameter is gone.

Actually: The two coexist, and which one applies depends on the exact model. Anthropic documents effort working alongside a token budget on one generation, while an older model in the same family still takes the token budget and rejects effort outright. There is also a separate advisory budget covering a whole agentic loop. Read the docs for the model you are calling.

Two Claude models, two different thinking APIs

Turn the dial down. Do not switch thinking off.

On the newest Claude models, disabling thinking is worse than lowering it. Anthropic reports that the model starts writing tool calls into visible text, where they silently never run, and leaks internal tags into the answer. The published guidance is to start at the default level, run your own evaluations, and move from there rather than reaching for off.

Try it

Filter the live index to models that reason. The separate reasoning-model category has all but dissolved. This step is an interactive widget; open the lesson to use it.

In one sentence

Effort is a word now, not a number, but the old number is still wired up on plenty of models. Check before you assume.