Reasoning Effort: Low, Medium, and High

Updated November 3, 2026 based on the xAI API reference: max_completion_tokens caps only the visible response; reasoning tokens do not count toward this limit, and effort controls their quantity. Source: Chat Completions. November 3 addendum: The lesson previously linked effort to the API (Chat Completions or Responses) and applied it to grok-4.20-0309-reasoning, which does not support it; according to xAI’s Reasoning page, it depends on the model (grok-4.7, grok-4.6, grok-4.5, grok-4.20-multi-agent). The example also read usage.completion_tokens, which is absent from Responses API responses (output_tokens, output_tokens_details.reasoning_tokens). Source: Reasoning.

Control Reasoning Depth

Reasoning effort lets you calibrate how much thinking the model invests before answering. It’s a direct lever for balancing quality, speed, and token consumption.

The model determines availability, not the API. According to xAI documentation:

Reasoning cannot be disabled on these models: without configuration, effort defaults to high. Only the parameter name changes depending on the API: reasoning_effort at the top level in Chat Completions, reasoning.effort inside a reasoning object with the Responses API.

{
  "model": "grok-4.7",
  "input": "Solve the system of equations: 3x + 2y = 12, x - y = 1",
  "reasoning": {
    "effort": "medium"
  }
}

Effort Levels

The documentation describes each level and its appropriate use:

• low: Uses some reasoning while staying fast. For latency-sensitive agentic uses and simple tool calls.
• medium: More reflection, for applications less sensitive to latency. For complex data analysis and reasoning over long contexts.
• high (default): Many more reasoning tokens. For very difficult problems, complex mathematics, multi-step logic.
• xhigh: Maximum depth, with corresponding latency. For the hardest problems, when quality matters more than response time.

xAI best practices add a simple rule: set effort explicitly on every request, and log it with every result, so your measurements remain comparable.

Reasoning models also reject three parameters: presence_penalty, frequency_penalty, and stop will fail the request.

Impact on Tokens

Effort level has a direct impact on reasoning token consumption:

• low: Few reasoning tokens, fast response, low cost
• medium: Moderate consumption, good compromise
• high and xhigh: Many reasoning tokens, slower response, high cost

These reasoning tokens are billed as output tokens (see Lesson 8). The output cap (max_completion_tokens in Chat Completions, max_output_tokens in the Responses API) limits only the visible response: reasoning does not count toward it, so effort controls their quantity.

Choosing the Right Level

Practical Example: Same Question, Three Efforts

for effort in ["low", "medium", "high"]:
    response = client.responses.create(
        model="grok-4.7",
        input="Explain why 0.1 + 0.2 != 0.3 in computing.",
        reasoning={"effort": effort},
    )
    usage = response.usage
    print(f"Effort {effort}: {usage.output_tokens} output tokens, "
          f"including {usage.output_tokens_details.reasoning_tokens} reasoning tokens")

The gap between levels depends on the question: measure it on your own cases rather than relying on a general ratio, and log the effort with every result.

Key Takeaways

• Effort depends on the model: low, medium, high (default), and xhigh on grok-4.7 and grok-4.6, without xhigh on grok-4.5; grok-4.20-0309-reasoning does not support it
• reasoning_effort in Chat Completions, reasoning.effort in the Responses API; reasoning cannot be disabled
• Effort level directly impacts the number of reasoning tokens and therefore cost
• Without configuration, effort defaults to high: set it explicitly on every request, as xAI recommends
• Adapt effort to the actual complexity of the task to optimize costs