xAI Grok

Grok models use the shared compatible endpoints — no dedicated interface: Chat Completions, Responses and Messages all work.

This page covers only what survives a trip through Amux — which parameters pass through, which get stripped, and which capabilities are reachable on which endpoint. For what the parameters mean upstream, see the xAI docs; we don't restate them.

Models#

Images can be base64 or a URL — with a URL the upstream fetches it, and returns invalid_image if it can't. Files go through the Responses API as input_file + file_url.

Parameter support#

Four states: passed through reaches the upstream unchanged; stripped is removed at the gateway and reported in the x-amux-dropped-params response header; ignored is sent but the upstream does nothing with it; rejected fails the request outright.

grok-build-0.1 still produces reasoning tokens — it just doesn't let you dial the effort.

Server-side tools#

⚠️ Available only on the Responses endpoint.

Passing server-side tools to Chat Completions or Messages neither errors nor works — the upstream accepts them and then ignores them, answering from the model's own knowledge. The workbench playground can't use them either.

Declare them like this:

{
  "model": "grok-4.6",
  "input": "What's the AI news today?",
  "tools": [{ "type": "web_search" }]
}

Availability and rates for each tool are defined by the model catalog; GET /v1/models returns them under pricing.tiers.

⚠️ The model decides how many times to call, and one request may call several times. A single "find me a recent post on X" triggered four x_search calls in testing. usage.server_side_tool_usage_details breaks down how many of each ran.

One search also pours several thousand tokens into the context, and those count towards usage too.

Using mcp:

{
  "model": "grok-4.6",
  "input": "Use deepwiki to tell me what facebook/react is",
  "tools": [
    { "type": "mcp", "server_label": "deepwiki", "server_url": "https://mcp.deepwiki.com/mcp" }
  ]
}

Each MCP call the model makes leaves an mcp_call item in output[], and the count lands in usage.server_side_tool_usage_details.mcp_calls.

grok-build-0.1 does not support the Responses endpoint, so it can't use server-side tools.

Usage fields#

Reasoning tokens are not inside completion_tokens. On Chat Completions, completion_tokens does not include reasoning_tokens; add them together for this call's output volume:

output = completion_tokens + completion_tokens_details.reasoning_tokens

The Responses endpoint's output_tokens already includes reasoning, so use it as-is.

Every call carries a fixed prompt overhead. Grok ships with a system prompt of its own — roughly 637 tokens on grok-4.6 and 495 on grok-4.5, most of which usually hits the cache.

Long context is tiered. Once a single request's input reaches 200,000 tokens it moves to the upper tier. The threshold is 200K on all four models — note that grok-build-0.1 tops out at 256K of context. The tiers themselves are defined by the model catalog.

Not supported#

service_tier and collections_search / file_search are not offered.

deferred and search_parameters are deprecated or inert upstream — passing them has no effect.