xAI Grok models

Choose the right model in the family

Start with the workload and service target, then pick the smallest model that passes a representative evaluation set.

• Choose Grok 4.6 for coding, knowledge work and tool-using agents that benefit from a large context window.
• Use low or medium reasoning for interactive work; evaluate high or xhigh only where the quality gain justifies latency and output cost.
• Prefer a smaller route for simple classification or extraction when your evaluation shows no benefit from the flagship model.

Inputs, outputs, context and tools

• Text and image input with text output through Responses and Chat Completions APIs.
• A 500K context window supports large codebases, long documents and extended agent history.
• Published tools include function calling, web search, X search and code execution.
• Reasoning effort supports low, medium, high and xhigh; high is the documented default.

Where this family fits

These are strong starting points—not universal recommendations. Test them with your own prompts, tools and acceptance criteria.

• Multi-file coding, debugging and repository analysis.
• Research agents combining web and X search with structured tools.
• Long-form knowledge work across large source collections.
• Multi-step workflows that need code execution and function calls.

Reasoning, context and cost controls

• Use prompt_cache_key or the Chat Completions conversation header to improve cache affinity on long agent loops.
• Compact or summarize context before the 500K limit instead of repeatedly sending an ever-growing transcript.
• Fresh internet or X information requires the corresponding search tool; the base model alone is not a realtime feed.

Send a minimal request

This example uses the first provider model ID in the table. If Model Plaza shows a different ID for your API key, copy the Model Plaza value exactly.

curl https://app.llmfly.ai/v1/chat/completions \
  -H "Authorization: Bearer $LLMFLY_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "grok-4.6",
    "messages": [{"role": "user", "content": "Explain this request in three bullets."}],
    "stream": false
  }'

Limits and compatibility notes

• Vendor pricing doubles for requests whose context reaches the published long-context threshold; confirm how the selected route bills it.
• Image input is limited to supported formats and size constraints; validate preprocessing before production.
• Newer Grok models do not support every legacy sampling option, including log probabilities.

Production evaluation checklist

• Run 20–50 representative tasks and define pass criteria before testing.
• Measure time to first token, total latency, input and output tokens, tool retries, and cost per successful task.
• Test the endpoint, streaming mode, tool schemas, and structured output used by the application.
• Verify 400, 401, 404, 429, and temporary 5xx handling with bounded retries.
• Keep the model ID in configuration and validate a fallback before moving production requests.

What to confirm in Model Plaza

Model Plaza shows which model IDs your API key can use, their billing multipliers, availability, supported endpoints, and compatibility notes. Confirm these details with a small request before deployment.

• Grok 4.6 uses a 500K context window.
• Reasoning effort can be configured for the workload.
• The LLMFly AI group must support the exact model name; unsupported names return 404.