TL;DR: Grok is the model family from xAI, the lab Elon Musk founded in 2023 that is now part of SpaceX. The current models are closed-weight and hosted: you reach them through an OpenAI-compatible API, through cloud catalogs such as Amazon Bedrock, Microsoft Foundry, and Oracle Cloud, or inside the Grok apps and X. The family's signature is live access to X and the web as built-in tools, backed by the X platform's data and distribution.
What it is and how you use it
xAI was founded in 2023 and has changed corporate shape twice since: it absorbed X, the social network, in an all-stock merger in March 2025, and in February 2026 it was itself acquired by SpaceX, under which it now operates as SpaceXAI. The Grok name, the grok.com apps, and the x.ai developer platform all survived the reshuffles, so for a developer the practical identity is unchanged - a single-vendor, closed-weight family served from the lab's own data centers and from partner clouds.
The family is best known for three things. The first is real-time grounding: the API exposes X search, web search, and code execution as server-side tools, so the model can read the live platform and the open web inside a single request - the design assumes the model's own knowledge is stale and the tools fill the gap. The second is distribution: Grok ships inside X, in standalone apps, and in Tesla vehicles, which puts it in front of an audience most labs reach only through partners. The third is its position at the frontier tier, where xAI leads each launch with the agentic suites its peers quote - Terminal-Bench and SWE-bench variants among them - as vendor claims; a multi-agent "Heavy" configuration, in which a lead model coordinates sub-agents, is offered on the API and in the top consumer tier. Alongside the text models sit the Grok Imagine image and video generators, voice models, and Grok Build, the lab's terminal coding agent, which launched in 2026 with MCP support and a plan mode.
On weights, xAI is a special case among the closed labs. It has published the weights of older generations - Grok-1 under Apache 2.0 in 2024, Grok 2 under a custom community license in 2025 - but with a lag of a year or more and under terms that are not open source in the OSI sense, and a promised Grok 3 release had not appeared by autumn 2026. Treat the current generation as closed: there is nothing to download, and nothing to self-host. One more fact an encyclopedia should not omit: the service has drawn regulatory attention more than once - an incident in July 2025 in which the X bot posted antisemitic content prompted an apology from xAI, and in January 2026 its image tools' generation of sexualized deepfakes led Indonesia and Malaysia to block the service temporarily and the UK and EU regulators to open formal investigations.
Access is entirely hosted. The first-party REST API is OpenAI-compatible, with Chat Completions and Responses endpoints, batch processing, prompt caching, structured outputs, and the server-side tools above, plus an official Python SDK and a provider for the Vercel AI SDK. The same models are in the cloud catalogs: Amazon Bedrock added the family in mid-2026 and tracks new releases within weeks, Microsoft Foundry has carried it since 2025, and Oracle Cloud lists the current flagship. Where it fits: products that need live social or web context, teams already standardized on X's ecosystem or on one of those clouds, and anyone running a frontier-tier bake-off who wants a fourth closed option beside GPT, Claude, and Gemini. The counterweights are single-vendor dependence and the governance scrutiny above, both of which belong in the evaluation.
Where it sits in the AI stack
Grok occupies the model layer as a hosted, closed family: your application calls the xAI API or a cloud catalog, the model optionally calls its live-search tools, and the response streams back:
Key tools and implementations
-
xAI API
The OpenAI-compatible first-party API with Chat Completions and Responses endpoints, batch, prompt caching, structured outputs, and built-in X search, web search, and code execution tools.
-
Cloud catalogs
The same models through Amazon Bedrock, Microsoft Foundry, and Oracle Cloud Infrastructure - billed inside your existing cloud account.
-
SDKs and Grok Build
An official Python SDK and a Vercel AI SDK provider, plus Grok Build, the lab's terminal coding agent with MCP support.
-
Grok apps, X, and older open weights
The consumer surfaces - grok.com, the mobile apps, X, and Tesla vehicles - and, for researchers, the Grok-1 and Grok 2 weights published on Hugging Face.
Related entries
- GPT OpenAI's family of closed-weight language models, served through its API, Microsoft Foundry, Amazon Bedrock, and the ChatGPT products.
- Gemini Google's multimodal model family, accessed through the Gemini API, Google Cloud, and Google's own products.
- Claude Anthropic's family of language models, accessed through an API and major cloud platforms rather than downloadable weights.
- Tool use (function calling) Letting a language model request actions by emitting structured calls that your code executes and whose results feed back in.
- Choosing a model A framework for weighing capability, latency, cost, privacy, and openness when picking a model for a task.