Picking a model is one lever (Tip #46). There’s a second one: how hard the model thinks. Most modern models expose a reasoning-effort setting — low → medium → high → xhigh/max — trading latency and cost for depth. (Some even have a none setting for extra speed.)
Rough guide, straight from OpenAI’s and Anthropic’s docs:
- low — execution-oriented coding, simple edits, drafting
- medium — everyday agentic coding (a solid default)
- high / xhigh — hard debugging, deep planning, security and code review
The counterintuitive part: reasoning effort can matter more than model size. A mid-tier model thinking hard can beat a bigger model thinking little. Anthropic says it plainly: “Tuning effort is often a better lever than switching models.”
Caveat: thinking isn’t free — and at high effort, models might over-engineer, adding refactors and abstractions you didn’t ask for. It also takes longer.
Where’s the dial? In Copilot, open /model and use ←/→ to set the reasoning effort for the highlighted model. Claude Code has /effort.
💡 Try this: A/B test it on one model. Set the lowest reasoning effort, give it a task, notice the outcome. /rewind to reset, then set the same model to the highest reasoning level and give it the same task. Do the results differ? Did it take longer?
🔗 Anthropic effort guide
🔗 OpenAI reasoning
Have you tuned reasoning effort?
🟢 Yes, it’s part of my workflow
🟡 Tried it once or twice
🔴 Haven’t thought much about it (my reasoning effort was low 😁)
⚪ I am going to try it now!
What did you observe from the 💡 Try this-experiment?