How to enable reasoning for Grok 4 Fast and reasoning effort for GPT 5 (OpenRouter)

General
1

Hi!

By default, the OpenRouter’s API makes GROK 4 Fast non-reasoning, and we can provide a enabled parameter to request if we wish to enable reasoning. We can also provide effort.
But I can’t figure out how to provide those OpenRouter additional params to the final API call. My crew uses the standard CrewAI LLM class.

And the second question. How do I change the GPT 5 reasoning effort?

Thanks in advance!

2

In theory, you should do the following to enable (and adjust the level of) reasoning for Grok 4 Fast on OpenRouter:

from crewai import LLM
import os

os.environ["OPENROUTER_API_KEY"] = ""

llm = LLM(
    model="openrouter/x-ai/grok-4-fast",
    temperature=0.7,
    reasoning_effort="high"  # or "low" or "medium"
)

print(
    llm.call(
        "Is the number 0.999... (with the 9s repeating forever) equal to 1, "
        "or is it less than 1? Justify your answer."
    )
)

The problem is, to my surprise, it seems like Grok 4 Fast is just too good at reasoning, even without the reasoning_effort parameter. So, I asked this question without passing that parameter. And, upon checking the OpenRouter logs, look what it says:

Tokens: 191 prompt → 666 completion, incl. 324 reasoning

See that? Apparently, it included reasoning tokens even with reasoning turned off.

2 Likes
3

Thanks a lot! Looking at my logs, I can see that you’re right. Perhaps OpenRouter has made some sort of mistake on their side.

1 Like
4

You can configure reasoning for both models directly in the CrewAI LLM class. For Grok 4 Fast, use the reasoning_effort parameter (“low”, “medium”, or “high”) to enable and control it . For GPT-5, the same reasoning_effort parameter should be used to adjust the reasoning level .