Grok

Use the exact model ID returned by GET /v1/models; no aliases or bare names are registered.

curl https://api.hop-base.com/v1/chat/completions \
  -H "Authorization: Bearer sk-your-key" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "grok-4.6",
    "messages": [{
      "role": "user",
      "content": "Summarize this quarter'\''s risks in five bullets."
    }]
  }'

Chat and reasoning models

Both POST /v1/chat/completions and POST /v1/responses are available to any OpenAI SDK client. Codex CLI cannot use Grok directly: it sends a namespace tool type that the model service rejects with 422 before generating anything. On /v1/responses, previous_response_id is removed and store is forced to false, so send the full input every turn.

422 Failed to deserialize the JSON body into the target type: tools[<n>].type: unknown variant `namespace`

Chat and image models are served by the Grok Suite group; video is served by Grok Video. A key reaches the video model only if it also includes Grok Video; one key can include both groups. Check GET /v1/models with the key you intend to use.

Request parameters

Applies to both endpoints. The gateway only rejects empty messages on Chat Completions and sets no bound on the parameters below; limits marked "official" come from xAI's reasoning guide and API reference.

xAI offers reasoning_effort on two models only:

• grok-4.6: low / medium / high / xhigh.
• grok-4.5: low / medium / high.
• grok-4.20-multi-agent-0309 uses reasoning.effort (reasoning_effort on Chat Completions), which sets how many agents collaborate.

presence_penalty, frequency_penalty, and stop are not supported (official); reasoning models reject them with an error. On grok-4.20 and newer, logprobs and top_logprobs are not supported either and token log probabilities are not returned, so omit them.

For every field, see API reference · Chat Completions and pick the Grok model family.

Billing and usage

Long-context requests

A request whose prompt exceeds 200,000 tokens is treated as a long-context request.

• The threshold counts non-cached input plus cached input. Output tokens do not count toward it.
• A request that lands exactly on 200,000 tokens does not cross it; it takes one token more.
• Affected requests are flagged long_context on the Usage page in the console, so you can tell them apart.

Reasoning tokens

Grok is a reasoning family, and the two endpoints report reasoning differently. This matters when you read usage programmatically.

Chat Completions reports reasoning separately in completion_tokens_details.reasoning_tokens. HopBase passes xAI's response body through unchanged, so a single reply can report completion_tokens: 1 with reasoning_tokens: 158.

Add the two together to get the real output length; on /v1/responses it is already included, so do not add it again.

Server-side tools

Grok can call built-in server-side tools during a request: web search, X search, code interpreter, document search, and file search. MCP calls reach your own remote service.

The number of tool calls a request made is shown on the Usage page in the console, broken down by tool type. See the signed-in Model Plaza for your effective rates.

Image models

The Grok image models grok-imagine-image, grok-imagine-image-2.0, and grok-imagine-image-quality use the standard OpenAI Images API and are synchronous only. Use POST /v1/images/generations for text to image and POST /v1/images/edits for image to image.

Shape comes from aspect_ratio and the pixel tier from resolution; there is no size. For the full parameters, measured output sizes, and what you are billed for, see Grok Imagine images and video.

Grok video (grok-imagine-video-1.5) belongs to the Grok Video group and uses the asynchronous task API; it is covered on the same page.

Next steps

• Grok Imagine images and video
• Codex CLI — direct connection over Responses
• Error codes and retries
• Concurrency, timeouts, and billing