GitHub added a reasoning level control to Copilot cloud agents on 3 August.
You now choose how hard the agent thinks. That is one setting with two effects that pull against each other: higher reasoning costs more and takes longer, lower reasoning is cheap and misses things.
This is the same idea Cursor shipped in July
Cursor's Router, released 22 July, routes each request to an appropriate model by cost and difficulty. Cursor's ₹649 India plan, on 28 July, is explicitly capped at "fixed medium effort".
Reasoning effort has become a first-class product control across the category in about six weeks. That is a sign the industry has accepted that maximum reasoning on every request is neither affordable nor necessary.
How to actually set it
A rough rule that holds up in practice:
- Low for mechanical work — renames, formatting, mass find-and-replace, dependency bumps
- Medium as the default for ordinary feature work, which is most of what an agent does
- High for anything where being wrong is expensive: migrations, auth, payments, data deletion, concurrency
Why the timing matters
OpenAI cut GPT-5.6 Luna by eighty percent on 30 July. Cheap inference plus an effort dial means the economics of running agents changed twice in five days.
If you costed out cloud agents before the end of July, that number is stale.