Grok API: set reasoning effort explicitly and leave sampling near 0.7/0.95

xAI recommends temperature 0.7 and top_p 0.95, explicit reasoning effort on every request, streaming, and long client timeouts.

Published October 10, 2026

From xAI’s own recommended settings for Grok reasoning models:

Setting	Recommendation
Sampling	`temperature: 0.7`, `top_p: 0.95`; quality usually drops when you stray
Reasoning effort	Set it explicitly every request; defaults differ by model
Streaming	Prefer it, so long generations don’t drop
Timeouts	At least 900 seconds per request
`max_output_tokens`	Above your longest expected answer (it caps only the visible answer)

Also log usage on every response. Cached tokens and reasoning tokens tell you whether your cache and effort settings are doing what you think.

Related tips

On xAI's Responses API, append every output item (including encrypted reasoning) to the next request. Dropping it hurts quality and cache hits.

Oct 16, 2026

reasoning_effort controls how hard the model thinks; verbosity controls how long the answer is. Adjust them instead of rewriting prompts.

Oct 13, 2026

OpenAI recommends versioning prompts in your repo, using typed variables, reviewing them in PRs and running evals on every change.

Oct 9, 2026

Sources