reasoning_effort is the most useful dial on Kimi K3 and the one most people never touch — which is expensive, because it defaults to max, the highest of the three levels. Unlike most models, K3 does not let you turn reasoning off at all. Effort is your only lever.
The three levels
| Level | Use it for |
|---|---|
low | Mechanical work: reformatting, renaming, extraction, boilerplate, narrow lookups |
high | The sensible working default. Most well-specified real work |
max | The API default. Hard reasoning: debugging, cross-document analysis, long agent runs |
There is no medium, and there is no off. Three levels is a coarser ladder than some competitors offer — Claude's models expose five — so the jumps between settings are larger and worth measuring.
The default is the expensive one
This deserves emphasis because it is unusual. Most APIs default to a middle setting. K3 defaults to max.
If you deployed without setting reasoning_effort, every request you have made has been at the top of the ladder. For a workload that is mostly extraction and formatting, moving to low can cut spend substantially with no measurable quality change — reasoning tokens bill as output at $15.00 per million.
Set it explicitly. Even setting it to max deliberately is better than inheriting it, because then it is a decision you can revisit.
Higher effort is not always more expensive overall
The counter-intuitive part, and it is real on agentic work: a model that reasons harder up front often takes fewer turns, makes fewer wrong tool calls, and needs less correction. Total spend for the finished task can be lower at max than at low.
The relationship is not monotonic, which is exactly why the advice is to sweep rather than assume. Run your evaluation set at all three levels and compare cost per completed task, not cost per request. For single-shot work the ranking is usually the obvious one; for multi-step agent loops it frequently is not.
A starting policy
- Extraction, formatting, classification →
low - Code generation, drafting, summarisation →
high - Debugging, cross-document analysis, agent planning →
max - Anything you have not measured →
high, then measure
The one thing not to do is leave it unset and assume you are on a sensible middle setting. You are not.
More on Kimi K3
Start with the complete Kimi K3 guide for the overview, or go deeper:
Ready to go deeper?
Read the full Kimi K3 guideFrequently Asked Questions
What is the default reasoning_effort on Kimi K3?
`max` — the highest of the three available levels. This is unusual: most APIs default to a middle setting. If you have not set the parameter explicitly, you are paying for maximum reasoning on every request.
Can I turn reasoning off on Kimi K3?
No. K3 always has thinking mode enabled; there is no equivalent of disabling it. The lowest setting is `reasoning_effort: low`, which reduces reasoning depth but does not eliminate it. For genuinely trivial high-volume work, a smaller non-reasoning model is the better tool.
Does lower effort make Kimi K3 cheaper?
Per request, yes — reasoning tokens bill as output at $15 per million. Per completed task, not necessarily: on agentic work, higher effort often means fewer turns and less rework, so total spend can fall as effort rises. Measure cost per finished task.
