Module 6 · EffortVerified 2 Nov 2026
Thinking effort
Effort controls how long a reasoning model thinks before it answers. More thinking helps hard problems and wastes time on easy ones.
Thinking effort
Drag the slider
LowMediumHigh
A balanced amount of reasoning. A good default for everyday coding work.
Depth
Speed
AI credits
- Implementing a clearly specified function
- Writing tests for existing code
- Reviewing a small diff
Trivial edits (you pay for thinking you do not need) and deep design problems (it may under-think).
Add pagination to GET /orders. Follow the pattern in /customers and add tests.
Match the effort to the task
Tap a task
Pick a task to see a suggested level.
| Topic | Detail |
|---|---|
| Where to set it | Open the model picker in the chat input, select a reasoning model, then use the > arrow next to its name to reveal the effort levels. |
| It varies by model | Level names and defaults differ between models and providers. Non-reasoning models (for example GPT-4.1 or GPT-4o) show no effort option at all. |
| Adaptive thinking | Some models decide per request whether, and how much, to think. VS Code enables adaptive reasoning where supported, so the default is often right. |
| It is remembered | Effort and context size are stored per session, and new chats start from the level you last used for that model. |
| Thinking tokens count | They are not shown in the reply but they use context window space and AI credits, so only raise effort for genuinely hard work. |
Do
- ✓ Start low, then raise effort if the answer misses
- ✓ Give high effort a clear question and evidence (logs, failing test)
- ✓ Ask for ranked hypotheses before any fix
- ✓ Drop back to low once the hard part is solved
Don't
- ✗ Leave high effort on for every request
- ✗ Use high effort to compensate for a vague prompt
- ✗ Expect effort to help a non-reasoning model
- ✗ Ignore latency and credit cost on routine edits
Where is the setting?
Open the model picker, choose a reasoning model and use the arrow next to it. Level names differ by model.