xAI
Very large context and a high reasoning ceiling — but almost no published prompting guidance for its text models, so treat advice about Grok with more caution than the rest.
How to prompt it
Structure with Markdown headings, not XML. XML tags are never mentioned anywhere in xAI’s documentation as a way to structure a prompt.
Write in the second person — “You are…” — and keep sections in a consistent order. This is the pattern xAI documents, though for its voice model rather than its text models.
Keep it short. xAI’s guidance for its newer models is that your prompt should be much shorter than before, and that short specific instructions are followed more reliably than long ones.
Put durable instructions, examples and reference documents at the very front — the documented reason is cache stability rather than attention, but the advice lands the same way.
Never mention a tool in the prompt that is not in the tool definition. The model follows instructions closely, and a mismatch produces bad answers.
Set reasoning effort deliberately. It defaults to high and cannot be disabled on the frontier models, so an easy task will otherwise cost you more than it needs to.
Strengths
- Large context across the whole line, and the flagship publishes no text output limit at all.
- Four graduated reasoning levels, higher than most competitors offer, including a multi-agent variant where the setting controls how many agents run rather than how deeply one thinks.
- First-class server-side search over the web and X, with domain allow and exclude lists. For anything recency-sensitive that is a real differentiator.
- Well-documented and aggressive prompt caching, with cached input priced at a fraction of full input.
Weaknesses
- Almost no official prompting documentation for the text models. The one structured guide xAI publishes is scoped to its voice model, and xAI does not claim it transfers. Anything you read about how to prompt Grok, here included, is inference from adjacent docs.
- Reasoning cannot be switched off on the frontier models and defaults to high, so trivial tasks carry a cost and latency floor. Penalty and stop parameters become hard errors alongside it.
- Its own documentation contradicts itself in at least two places, on the reasoning default and on the knowledge cutoff. That is a reliability signal in itself.
- No documented EU data residency, and no current-generation open weights — only two ageing releases, the more recent under a restrictive licence.
Models
xAI publishes no maximum output figure for any text model except the flagship, which it describes as having no text output limit. The dashes below are missing figures, not zeroes.
Grok 4.6
Frontier model for coding, agentic tasks and knowledge work.
- Context
- 500K
- Max output
- [object Object]
Good for: Hard reasoning and agentic work where you also want very long output.
Watch out: Reasoning cannot be turned off, and defaults to high.
Grok 4.3
Efficient long-context alternative.
- Context
- 1M
- Max output
- —
Good for: Very long documents at lower cost than the flagship.
Watch out: Prompts longer than 200k tokens move the whole request to a higher price tier.
Grok 4.20
Available as reasoning, non-reasoning and multi-agent variants.
- Context
- 1M
- Max output
- —
Good for: Choosing explicitly whether you want reasoning at all, which the frontier models do not allow.
Watch out: Log-probability options are silently ignored on this generation and newer.
Sources and grounding
- 1ModelsxAI / SpaceXAILineup, context windows, pricing tiers and the note that roles may appear in any order.
- 2ReasoningxAI / SpaceXAIReasoning effort levels, the high default, and that reasoning cannot be disabled.
- 3Prompt caching best practicesxAI / SpaceXAIThe front-load and never-edit-history guidance.