In six weeks, reasoning effort went from an implementation detail to a control you are expected to set.
Cursor shipped Router on 22 July to route by cost and difficulty, capped its ₹649 India plan at fixed medium effort on 28 July, and GitHub added a reasoning level to Copilot cloud agents on 3 August.
What the setting actually trades
Higher effort means more tokens spent reasoning before answering. That buys better handling of problems where the answer depends on holding several constraints in mind at once.
It costs money and wall-clock time. On a task where the answer is obvious, the extra reasoning changes nothing and you pay for it anyway.
A rule that holds up
- Low — mechanical transformations: renames, formatting, dependency bumps, mass find-and-replace, boilerplate
- Medium — ordinary feature work, which is most of what an agent is asked to do
- High — migrations, auth, payments, concurrency, data deletion, anything where a wrong answer is expensive to discover later
The mistake almost everyone makes first
Setting high everywhere. It sounds like the safe choice and it is not — you pay a premium on every trivial task, responses get slow enough that people stop using the agent, and the quality gain on simple work is roughly zero.
The opposite mistake is rarer but worse: leaving everything low to save money, then merging a migration nobody checked.
The economics moved under you
OpenAI cut GPT-5.6 Luna eighty percent on 30 July. Cheaper inference makes high effort more affordable than it was a month ago, which means the sensible default may have shifted up.
Re-run your own numbers rather than inheriting a rule of thumb from June.