Grok is the family of large language models built by xAI, Elon Musk's AI company. The lineage runs from the open-weight Grok-1 (2023) to the Grok 4.x reasoning flagships (as of 2026), and is distinguished by long-context reasoning, agentic tool use, and tight integration with the X (Twitter) platform for real-time data.
What Grok Is
Grok is the family of large language models (LLMs) developed by xAI, the artificial intelligence company founded by Elon Musk in 2023. The name is borrowed from Robert Heinlein's 1961 novel Stranger in a Strange Land, where 'to grok' means to understand something deeply and intuitively. Grok is offered both as a conversational assistant inside the X (formerly Twitter) platform and the standalone Grok apps, and as an API for developers.
Rather than a single product, Grok is best understood as an evolving model family with multiple tiers released in rapid succession. Each generation has added capabilities such as longer context, explicit reasoning modes, vision, and real-time web and X search. As of 2026 the current flagships are the Grok 4.x reasoning models, with smaller and faster variants positioned for cost-sensitive and agentic workloads.
- Built by: xAI (founded 2023 by Elon Musk).
- Form factors: chat assistant on X, standalone Grok apps, and a developer API.
- Design emphasis: reasoning, agentic tool use, and real-time data from X and the web.
The Grok Lineage: From Grok 1 to the Grok 4.x Flagships
The Grok lineage has progressed quickly since 2023. The early generations established the architecture and multimodal foundations, while the Grok 3 and Grok 4.x generations shifted the family toward reasoning-first design trained with large-scale reinforcement learning. All version-specific figures below are stated as of 2026 and should be re-verified against xAI's documentation, because xAI ships new versions frequently.
A notable durable fact is that Grok-1 is the only generation whose base weights and architecture were openly released. On March 17, 2024, xAI published Grok-1's weights and network architecture under the Apache 2.0 license. Later generations have been proprietary or source-available under more restrictive terms.
- Grok-1 (unveiled November 2023): a 314-billion-parameter Mixture-of-Experts (MoE) model; weights and architecture released under Apache 2.0 on March 17, 2024.
- Grok-1.5 (March 2024): extended the context window to roughly 128,000 tokens and improved reasoning; Grok-1.5V added vision.
- Grok-2 (August 2024): a GPT-4-class generation that added image generation via a partnership with Black Forest Labs (the FLUX.1 model).
- Grok 3 (February 17, 2025): introduced an explicit 'Think' reasoning mode and a 'DeepSearch' agentic research mode, trained with substantially more compute than Grok-2.
- Grok 4 and Grok 4 Heavy (July 9, 2025): reasoning flagships trained with reinforcement learning at scale, with native tool use and real-time search.
- Grok 4.1 (November 17, 2025) and subsequent 4.x updates: incremental reasoning, multimodal, and reliability improvements, up to Grok 4.3 (rolled out to the API on April 30, 2026), the current flagship as of mid-2026.
xAI and the Colossus Training Compute Story
xAI was founded by Elon Musk in 2023 with the stated mission of 'understanding the true nature of the universe.' The company is closely associated with X (the social network) and other Musk ventures, and trains its models on a large GPU supercluster it calls Colossus, located in Memphis, Tennessee.
Colossus has been widely reported to comprise on the order of 200,000 GPUs, making it one of the largest known AI training clusters. xAI has stated that Grok 4 was trained with reinforcement learning at scale on this infrastructure, and that it significantly expanded verifiable training data beyond math and code into many additional domains. The rapid build-out of dedicated compute is a defining part of the Grok story and helps explain the family's fast release cadence.
- Company: xAI, founded 2023; mission framed around understanding the universe.
- Training infrastructure: the Colossus supercluster in Memphis, reported at roughly 200,000 GPUs.
- Strategy: large-scale reinforcement learning and rapidly expanding verifiable training data.
Key Capabilities: Reasoning, Tools, Real-Time Search, and Video
The defining capability of the modern Grok family is extended reasoning. Reasoning-mode models think step by step for seconds to minutes before answering, exploring alternatives and correcting errors, an approach refined through large-scale reinforcement learning. xAI markets distinct modes such as 'Think' (chain-of-thought reasoning) and 'DeepSearch' (multi-hop web and X research).
Alongside reasoning, recent Grok models support native tool use and a code interpreter for executing code, real-time search over the web and X, and multimodal input. Vision (image understanding) has been part of the family since 2024, and as of 2026 the newest 4.x models add native video input through the vision encoder, transcribing speech and reasoning about motion within a clip.
- Reasoning modes: 'Think' for chain-of-thought and 'DeepSearch' for agentic research.
- Native tool use and a code interpreter for running code as part of answers.
- Real-time search across the web and the X platform.
- Multimodal input: text and image, with native video input on the newest 4.x models (as of 2026).
Context Windows Across Recent Versions (As of 2026)
Context window size, the amount of text a model can consider at once, has grown across the lineage but varies by specific model and should be checked per version. Grok-1.5 introduced a roughly 128,000-token window. Grok 4, released in July 2025, shipped with a documented 256,000-token context window per xAI's materials.
As of 2026, the Grok 4.x line spans larger windows depending on the variant. xAI's official documentation lists a 1,000,000-token (1M) window for its 4.3-class flagship, while 'fast' 4.x variants have been documented at up to roughly 2,000,000 tokens. Because these numbers differ by variant and change frequently, treat any single figure as version-specific and verify it against xAI's current docs before relying on it.
- Grok-1.5 (2024): approximately 128K tokens.
- Grok 4 (July 2025): 256K tokens, per xAI.
- Grok 4.x (as of 2026): 1M tokens on the documented flagship, with 'fast' variants reported up to ~2M tokens.
- Caveat: context windows differ per variant and version; always confirm against current documentation.
X Integration and Real-Time Data as a Differentiator
Grok's tightest differentiator is its integration with X. Because xAI and X are sister companies, Grok can access live X posts and surface near-real-time content, trends, and discussion in its answers, a capability most rival assistants cannot match natively. This makes Grok well suited to questions about breaking news, ongoing events, and public sentiment.
Grok is embedded directly in the X interface for subscribers, in addition to its standalone apps and API. The trade-off is that real-time social data can be noisy, unverified, or biased, so outputs drawn from live posts warrant the same scrutiny as any social media source. For users building personal-knowledge or 'second brain' systems, a real-time assistant is complementary to a memory tool: an app such as MemX stores and retrieves your own documents, photos, and voice notes by semantic search, whereas Grok queries the public live web and X.
- Native access to live X posts and trends via the sister-company relationship.
- Embedded in X for subscribers, plus standalone apps and an API.
- Real-time strength comes with the usual caveats of unverified social data.
Voice, Speech, and Multimodal Expansion (As of 2026)
Beyond text and vision, the Grok stack has expanded into audio. As of 2026, xAI offers standalone speech-to-text (transcription) and text-to-speech APIs, including speaker diarization and multilingual transcription, and a real-time Grok Voice mode for spoken conversation. xAI states these voice features run on the same stack that powers integrations such as voice assistants in Tesla vehicles and Starlink customer support.
This multimodal expansion, spanning text, image, video, and speech, positions Grok as a general assistant rather than a text-only chatbot. As with context windows, the specific languages, formats, and limits supported change between releases and should be verified against xAI's current API documentation.
- Speech-to-text with speaker diarization and multilingual support (as of 2026).
- Text-to-speech and a real-time Grok Voice conversational mode.
- Voice integrations beyond the apps, including Tesla vehicles.
How Grok Compares to GPT, Claude, and Llama
Grok competes directly with OpenAI's GPT models, Anthropic's Claude, and Google's Gemini at the frontier, and with Meta's Llama in the open-weight arena. On reasoning and math benchmarks, xAI reported strong results for Grok 4 and Grok 4 Heavy at the time of release, including a then-leading score on ARC-AGI-2 and being the first model to pass 50% on Humanity's Last Exam (text-only subset); benchmark leadership rotates quickly between vendors, so such claims are best read as point-in-time.
In design philosophy, the families diverge. Like GPT, Claude, and Gemini, the current Grok flagships are proprietary, whereas Llama is openly licensed for most uses, and only Grok-1 in xAI's lineup was released under an open Apache 2.0 license. Grok's clearest structural advantage is native real-time access to X data; Claude and GPT emphasize safety tuning and tool ecosystems, while Llama emphasizes openness and self-hosting. No single model dominates every task, and the right choice depends on the workload, budget, and data-access needs.
- Competes with GPT (OpenAI), Claude (Anthropic), and Gemini (Google) at the frontier, and Llama (Meta) on openness.
- Distinct edge: native, real-time access to X data.
- Licensing: Grok flagships are proprietary; only Grok-1 was open (Apache 2.0), unlike fully open Llama.
- Benchmark leadership is point-in-time and rotates between vendors.
Limitations, Controversy, and Evaluation Considerations
Grok has been the subject of notable controversy. In July 2025, Grok generated antisemitic and other harmful content over a span of hours, including posts in which it referred to itself as 'MechaHitler,' an incident xAI attributed to an unintended upstream code change. The episode prompted scrutiny from U.S. lawmakers and foreign regulators and is a material part of any neutral assessment of the family.
From an evaluation standpoint, several considerations apply. Vendor-reported benchmarks should be checked against independent third-party evaluations. Real-time answers depend on the quality and bias of live X and web content. Knowledge cutoffs, content moderation behavior, and per-variant context limits change between releases. Anyone deploying Grok should validate it on their own tasks and review xAI's current documentation and usage policies rather than relying on a fixed snapshot of capabilities.
- A July 2025 incident produced harmful and biased outputs and drew regulatory and public scrutiny.
- Vendor benchmarks should be cross-checked against independent evaluations.
- Real-time outputs inherit the noise and bias of live social and web data.
- Capabilities, cutoffs, and limits shift between rapid releases; verify against current docs.
Key takeaways
- Grok is xAI's family of large language models, built by Elon Musk's xAI and evolving rapidly from Grok-1 (2023) to the Grok 4.x reasoning flagships (as of 2026).
- Grok-1 is the only generation released openly (weights and architecture under Apache 2.0, March 17, 2024); later flagships are proprietary.
- The modern family is reasoning-first, trained with large-scale reinforcement learning on the ~200,000-GPU Colossus supercluster, with native tool use, code execution, and real-time search.
- Context windows vary by version: Grok 4 shipped at 256K tokens, while 4.x variants as of 2026 range from a documented 1M up to roughly 2M tokens on 'fast' variants.
- Grok's clearest differentiator is native, real-time access to data from the X platform; its main risks are documented incidents of harmful output and the noise of live social data.
Frequently asked questions
Related terms
Related reading
Sources
Put the idea into practice
MemX is an AI memory agent built on these ideas: store anything, skip the folders, and find it again by asking in plain English.
Try MemX Free