Knowledge tree
Offensive Security
Shells
File Transfer
Web Applications
Reasoning
Some models can spend more work on a task before producing the final answer. APIs and agent harnesses often expose this choice as a reasoning effort level such as low, medium, or high.
The setting does not describe the model’s internal process step by step. In practical terms, it asks the model to spend more or less effort on the request. A higher level can help when the task involves connecting several pieces of evidence, testing hypotheses, or revising a decision after new results arrive.
When it helps
Extracting one value from JSON or classifying a straightforward input usually needs little reasoning. Investigating an ambiguous failing test can be different: the model may need to inspect several files, understand a recent change, run the test, and interpret the next error.
More effort may help with the second task. It does not guarantee a better answer. The model can also explore unnecessary paths or take longer to reach a simple conclusion.
The useful way to choose a level is to test it on representative work. If medium consistently handles the tasks that matter, always selecting high merely because it is available may add no value.
Time, tokens, and cost
Additional reasoning commonly increases latency and usage. Accounting differs by API: a provider may include reasoning tokens in output usage, expose them as a separate detail, or report them in another form.
Measure the complete result:
- Whether the task was completed correctly.
- How long it took.
- How many tokens or calls it consumed.
The pricing side is covered in API Pricing. A small increase may not matter for one request, but an agent can pay it again on every inference in a long run.
Reasoning between tool calls
In an agent, reasoning does not happen only before the final answer. The model can decide to read a file, inspect the result, change its hypothesis, and run a test. The test output then affects the next decision.
For example, a coding agent may begin by suspecting the implementation, read the test, and discover that the expected value is stale. Its next action changes because new evidence arrived. That result-dependent sequence is part of the agent loop.
Higher effort can affect both analysis and the tool calls the model chooses. The useful question is not whether it produces a longer visible explanation, but whether the setting helps complete the task more reliably.