Grok is a family of large language models and the conversational assistant built on them, developed by the company xAI and first released in November 2023. Among AI models & assistants, Grok is distinguished by its tight integration with the X social platform, its access to real-time posts and web search, and a deliberately less restrained conversational style than most competing assistants.
The name, sometimes written xAI Grok, refers both to the underlying model generations, from Grok-1 to Grok 4 and later variants, and to the consumer product available inside X, as standalone mobile and web apps, and through a developer API.
How Grok is built and trained
Grok models are decoder-only transformer language models trained on large text corpora and then refined with supervised fine-tuning and reinforcement learning. The first generation, Grok-1, was a mixture-of-experts model with about 314 billion parameters, of which only a fraction is active for each token, which reduces inference cost relative to a dense model of the same size.
Later generations were trained on xAI’s Colossus supercomputer in Memphis, a cluster built with very large numbers of NVIDIA GPUs. xAI has published fewer architectural details for recent models than for Grok-1, so claims about exact parameter counts or training data composition for newer versions should be treated as unconfirmed unless officially stated.
The training emphasis shifted over time from general chat toward reasoning through reinforcement learning. Grok 3 introduced a visible reasoning mode, and Grok 4 was presented as having substantially more compute devoted to reinforcement learning, together with native tool use such as search and code execution.
Grok model generations
The Grok family has evolved in rapid release cycles, with major versions appearing roughly every six to twelve months and smaller variants in between. The table below outlines the main generations and what each one added.
| Generation | Released | Notable additions |
|---|---|---|
| Grok-1 | November 2023 | Mixture-of-experts base model; weights released openly in March 2024 |
| Grok-1.5 and 1.5V | Spring 2024 | Context window of 128,000 tokens; first image understanding |
| Grok-2 | August 2024 | Stronger reasoning and coding; built-in image generation |
| Grok 3 | February 2025 | Think reasoning mode and DeepSearch research agent |
| Grok 4 and Grok 4 Heavy | July 2025 | Heavier reinforcement learning; multi-agent Heavy variant |
| Grok 4 Fast and Grok Code Fast | Second half of 2025 | Lower-cost variants for speed and agentic coding |
xAI has also released open weights for older models, starting with Grok-1 under the Apache 2.0 license and later Grok 2.5, while keeping its current frontier models proprietary. This places Grok between fully closed families and openly licensed families such as Llama or gpt-oss.
Product features and access
The Grok assistant combines the language model with real-time retrieval, multimodal input and generation tools. Its most distinctive feature is access to public posts on X, which lets it summarize breaking discussions faster than assistants that rely only on web search indexes.
| Feature | What it does | Typical use |
|---|---|---|
| Real-time X search | Reads and summarizes recent public posts | Tracking news and public reactions |
| DeepSearch | Runs multi-step web and X research | Background briefings and comparisons |
| Think mode | Spends more computation on reasoning before answering | Mathematics, logic and coding problems |
| Image generation and editing | Creates and modifies images from prompts | Illustrations and social media visuals |
| Voice mode | Spoken conversation in real time | Hands-free questions and practice |
| Developer API | Programmatic access to Grok models | Building chatbots and agents |
Access is tiered. Basic usage is available with limits on free accounts, while higher limits, the newest models and heavier reasoning variants require paid X Premium tiers or dedicated SuperGrok subscriptions. Developers pay per million input and output tokens through the xAI API, which follows request formats compatible with common industry conventions.
How Grok compares with other assistants
Grok competes directly with general-purpose assistants such as ChatGPT, Claude and Gemini, and the main differences lie in data access, tone and content policy rather than in basic capability. On public reasoning and coding benchmarks, recent Grok versions have scored among the leading frontier models, although benchmark rankings change with each release.
| Aspect | Grok | Typical competing assistants |
|---|---|---|
| Real-time social data | Native access to public X posts | Web search without a dedicated social feed |
| Default tone | Informal, humorous, fewer refusals | More neutral and cautious by default |
| Content moderation | Looser limits on edgy or political topics | Stricter policies on sensitive content |
| Open weights | Older generations released | Varies from fully closed to fully open |
| Ecosystem integration | X platform and xAI apps | Office suites, search engines or developer tools |
The less restrictive style appeals to users who find other assistants overly cautious, but it also increases the risk of offensive or misleading outputs, which is the central trade-off of the product’s positioning.
Limitations, controversies and responsible use
Like every large language model, Grok can hallucinate facts, misread sources and present uncertain claims confidently. Its reliance on X posts adds a specific risk: social media content is fast but often unverified, so summaries of developing events can amplify rumors or misinformation before corrections appear.
Grok has also been the subject of public incidents involving harmful outputs. In 2025, xAI acknowledged that unauthorized changes to system prompts had caused the assistant to insert unrelated political claims into answers, and a later update produced antisemitic responses that the company removed and apologized for. These events drew attention to system prompt governance and the effect of loosened safety settings.
- Verify time-sensitive claims against primary sources rather than relying on summaries of posts.
- Avoid sharing confidential data unless the organization has reviewed xAI’s data retention and training terms.
- Review generated images for likeness, copyright and misuse risks before publication.
- Expect behavior changes between versions, because system prompts and models are updated frequently.
Understanding it simply
Grok is an AI chatbot made by the company xAI. You type or speak a question, and it writes an answer, explains a topic, helps with code or creates a picture. What sets it apart is that it can read what people are posting on X right now, so it can tell you what everyone is talking about today.
It is a bit like a well-read friend who also follows the news feed all day. That makes it quick and current, but a friend repeating the latest posts can also repeat mistakes, so important facts are still worth checking elsewhere.
Applied in practice
A communications team at a mid-sized airline monitors public reaction during a major winter storm that forces the cancellation of about 300 flights in one day. The team needs to know within minutes which complaints are spreading and which rumors about refunds are circulating.
An analyst asks Grok to summarize public X posts mentioning the airline in the last two hours, grouped by theme, and to flag claims about compensation. Within a minute, Grok returns five recurring themes, including confusion over rebooking links and a false rumor that all refunds were suspended, along with representative posts for each theme.
The team verifies the refund rumor against its own policy, publishes a clarification and updates its rebooking page. Grok shortened the first assessment from roughly an hour of manual scrolling to a few minutes, while human staff kept responsibility for checking every claim before responding publicly.