xAI Grok API

Grok models have direct access to live X platform data streams, enabling the model to retrieve and incorporate current tweets, trends, and social discourse into generation tasks without requiring separate API calls or external data fetching. This is implemented via server-side integration with X's data infrastructure, allowing the model to reference real-time events and conversations during inference rather than relying on training data cutoffs.

Grok-2 is exposed via an OpenAI-compatible REST API endpoint, allowing developers to use standard OpenAI client libraries (Python, Node.js, etc.) with minimal code changes. The API implements the same request/response schema as OpenAI's Chat Completions endpoint, including support for system prompts, temperature, max_tokens, and streaming responses, enabling drop-in replacement of OpenAI models in existing applications.

Grok-Vision extends the base Grok-2 model with vision capabilities, accepting images as input alongside text prompts and generating text descriptions, analysis, or answers about image content. Images are encoded as base64 or URLs and passed in the messages array using the 'image_url' content type, following OpenAI's multimodal message format. The model processes visual and textual context jointly to answer questions, describe scenes, read text in images, or perform visual reasoning tasks.

The API supports Server-Sent Events (SSE) streaming via the 'stream: true' parameter, returning tokens incrementally as they are generated rather than waiting for the full completion. Each streamed chunk contains a delta object with partial text, allowing applications to display real-time output, implement progressive rendering, or cancel requests mid-generation. This follows OpenAI's streaming format exactly, with 'data: [JSON]' lines terminated by 'data: [DONE]'.

The API supports OpenAI-style function calling via the 'tools' parameter, where developers define a JSON schema for available functions and the model decides when to invoke them. The model returns a 'tool_calls' response containing function name, arguments, and a call ID. Developers then execute the function and return results via a 'tool' role message, enabling multi-turn agentic workflows. This follows OpenAI's function calling protocol, supporting parallel tool calls and automatic retry logic.

The API provides a fixed context window size (typically 128K tokens for Grok-2) and supports token counting via the 'messages' parameter to help developers manage context efficiently. Developers can estimate token usage before sending requests to avoid exceeding limits, and the API returns 'usage' metadata in responses showing prompt_tokens, completion_tokens, and total_tokens. This enables sliding-window context management, where older messages are dropped to stay within limits while preserving recent conversation history.

The API exposes standard sampling parameters (temperature, top_p, top_k, frequency_penalty, presence_penalty) that control the randomness and diversity of generated text. Temperature scales logits before sampling (0 = deterministic, 2 = maximum randomness), top_p implements nucleus sampling to limit the cumulative probability of token choices, and penalty parameters reduce repetition. These parameters are passed in the request body and affect the probability distribution during token generation, enabling fine-grained control over output characteristics.

The xAI API supports batch processing mode (if available in the pricing tier), where developers submit multiple requests in a single batch file and receive results asynchronously at a discounted rate. Batch requests are queued and processed during off-peak hours, trading latency for cost savings. This is useful for non-time-sensitive tasks like data processing, content generation, or model evaluation where 24-hour turnaround is acceptable.

Grok API exposes multiple model variants (Grok-2, Grok-Vision, etc.) with explicit versioning, allowing developers to select models based on capability requirements and cost-performance tradeoffs. The API supports model aliasing (e.g., 'grok-latest' routes to the current production version), enabling automatic upgrades without code changes. Developers can specify exact versions for reproducibility or use floating versions for automatic optimization.

Grok API returns structured error responses with specific error codes (rate_limit_exceeded, invalid_request_error, etc.) and includes retry-after headers for rate limit errors. The API is designed to be compatible with OpenAI's error handling patterns, allowing developers to reuse existing retry logic. SDKs provide built-in exponential backoff and jitter to handle transient failures gracefully.