Every model has an Effort level: Low, Medium, or High. Effort sets how long the model works on a task before it settles on an answer. More effort means deeper answers on hard tasks, with higher token use and more time; less effort means faster, leaner responses.
Effort is a behavioral signal, not a strict token cap. At lower effort the model still works harder on genuinely difficult problems, it just does less than it would at a higher level for the same problem. The practical rule: start at the lowest level that meets your quality bar, and step up only when the answer needs more depth.
Deep research and multi-step agentic runs at higher effort are exactly the kind of work the usage is meant to cover, so leaving everything on High spends your budget faster than most tasks require.
Effort levels
| Level | What it does | Best for |
|---|---|---|
| Low | Fastest and most token-efficient. Fewest tokens, quickest turnaround, scoped to what was asked. | Lookups, short rewrites, translations, formatting, simple classification, high-volume tasks. |
| Medium | Balanced depth and efficiency. A strong default for everyday work. | Most day-to-day drafting, summaries, tool-assisted tasks, general Q&A. |
| High | Deepest reasoning and the most thorough analysis. Highest token use and latency. | Complex reasoning, planning, tricky debugging, nuanced analysis where quality matters most. |
Admins can set the default Effort level (and Thinking behavior) for the tenant in the Console. Individual users can still adjust Effort per model for a given task unless restricted.
The Thinking toggle
Some models also have a Thinking toggle (On/Off). Thinking controls whether the model reasons through the problem step by step before it starts executing, rather than answering immediately.
- Thinking Off: The model responds directly. Faster and more efficient. Good for straightforward requests where the path to the answer is obvious.
- Thinking On: The model reasons first, then answers. This improves reliability on multi-step problems, math, planning, and debugging, with additional tokens and time.
Effort and Thinking are related but separate: Effort sets how much total work goes into the whole response, while Thinking decides whether the model reasons before acting. Some models expose only Effort, some expose both. On models that have both, pairing Thinking On with Medium or High effort gives the most thorough results, and Thinking Off with Low effort gives the leanest.
How effort shows up with tools and agents
When a model uses tools or runs as an agent, Effort affects the tool calls themselves, not just the written answer:
- Lower effort tends to combine steps into fewer tool calls, skip preamble, and go straight to the action with a short confirmation.
- Higher effort tends to make more tool calls, explain its plan first, and give fuller summaries of what it did.
This is why Effort is a strong lever for long agentic runs: lower effort is not only fewer output tokens, it is also fewer round trips.
Choosing the right setting
A quick decision guide:
| If your task is... | Effort | Thinking |
|---|---|---|
| A quick lookup, translation, or format fix | Low | Off |
| Everyday writing, summarizing, or routine tool use | Medium | Off or On |
| Multi-step reasoning, planning, or code debugging | High | On |
| A high-volume or latency-sensitive workflow | Low | Off |
| Deep research across many sources | Medium to High | On |
Two habits that get the most value:
- Do not leave everything on High. High is built for hard problems. Most work is well served on Medium, and a lot is well served on Low.
- Match model and effort together. Effort tuning compounds with model choice. An efficient model (for example Kimi or GLM) on Medium effort often meets your quality bar at a fraction of the token burn of a frontier model on High.
How this affects your budget
Higher Effort and Thinking On increase token usage per message.
If you exhaust your usage, there are no lockouts: you revert to standard models and default settings and keep working. If your admin has enabled Extra Usage, higher effort levels and longer agentic runtimes continue by drawing from your Extra Usage allowance after your budget is empty.
Best practices
- Set effort explicitly. Do not rely on a single level for everything. The right starting point depends on the task and the model.
- Start low, step up. Begin at the lowest level that meets your quality bar and raise it only when the output needs more depth.
- Use Low for speed and volume. Simple, repetitive, or latency-sensitive tasks are well matched to Low.
- Reserve High and Thinking On for hard problems. Complex reasoning, planning, and debugging are where they pay off.
- Tune effort per workload. Pick a level that fits the task type and keep it consistent within a given piece of work.
- Combine with smart model selection. The most efficient model that clears your quality bar, set to the lowest effort that does the job, is the strongest combination.
Key takeaway
Effort and Thinking let you dial capability up or down on a single model. Keep both as low as your quality bar allows, raise them for genuinely hard tasks, and pair them with an efficient model to get near-frontier results while keeping your usage intact.