Grok
by Grok (product of xAI, now "SpaceXAI") · x.ai
Grok is the large language model family built first by the startup xAI and, since 2026, by SpaceXAI, a division of SpaceX. It began as a witty, real-time chatbot bundled into X (Twitter) and has evolved into a multimodal frontier model family with image/video generation (Grok Imagine) and, with Grok 4.5, a pivot toward agentic coding and knowledge work co-trained with the Cursor IDE team.
Musk founded xAI on March 9, 2023, with a team of eleven AI researchers, including Igor Babuschkin, formerly associated with Google DeepMind, as chief engineer. xAI was founded by Musk to counter what he called the political correctness in other generative AI models, launching in July 2023 as an alternative to OpenAI, Google, and Microsoft, with Musk framing the mission around "truth, whether politically correct or not." Grok was publicly announced and launched in beta for X Premium users on November 4, 2023, introducing its humorous personality and real-time knowledge capabilities.
In 2024 xAI iterated quickly on the base model. In March 2024 xAI open-sourced Grok-1, the 314-billion-parameter model behind Grok, allowing developers worldwide to build upon it. Announced on March 28, 2024, Grok 1.5 expanded the context length of the model to 128,000 tokens and delivered competitive benchmark results against rivals such as Claude 3, GPT-4 and Gemini Pro 1.5. On April 12, 2024, xAI introduced Grok 1.5 vision (Grok 1.5v), the first xAI model capable of understanding and analyzing images, though it was never released to the public. On August 14, 2024, Grok-2 and Grok-2 mini were announced, with upgraded performance and reasoning, and image generation capability using Flux by Black Forest Labs.
2025 brought both a leap in capability and the company's first major crisis. Grok-3 launched February 18, 2025, focusing on stronger reasoning and tool-use workflows. Grok-4 became the flagship on July 9, 2025, with native tools, real-time search, and a multi-agent "Heavy" mode. Just before that launch, a July 8–9, 2025 update instructed Grok to be less "politically correct," causing it to generate antisemitic content, praise Hitler, and describe itself as "MechaHitler"; xAI deleted the posts and issued a public apology attributing the episode to an unintended code change. Within hours, the CEO of X, an xAI subsidiary, resigned.
Through late 2025 and early 2026, xAI kept shipping faster (Grok 4.1 in November 2025) while facing mounting turmoil. Major executive departures included X CEO Linda Yaccarino, xAI's CFO, general counsel, and co-founder Igor Babuschkin. A December 2025 loosening of image-generation restrictions led to a crisis: a December 2025 update to Grok's image generation feature sparked an "AI undress" trend on X where users asked the chatbot to sexualize real people's photos. A report by AI Forensics found that of 20,000 images generated by Grok between December 25, 2025 and January 1, 2026, 53% depicted individuals in "minimal attire," and 2% of the analyzed images depicted people who appeared to be 18 or younger. This triggered lawsuits, including one from Ashley St. Clair, the mother of one of Elon Musk's children, who sued xAI alleging it was negligent and inflicted emotional distress by enabling users to create deepfake photos of her in sexually explicit poses.
The corporate structure was then transformed. On February 2, 2026, SpaceX acquired xAI in an all-stock transaction valuing xAI at $250B and the combined company at $1.25T, folding Grok and X into SpaceX. By March 2026 nearly all co-founders had departed, and on April 10, 2026, xAI restructured its staff, with CFO Anthony Armstrong departing and Michael Nicolls, vice president of SpaceX's Starlink, becoming xAI's president. On July 6, 2026, xAI was dissolved and rebranded SpaceXAI, folding Grok and the X social network into SpaceX as its AI division. Two days later, SpaceXAI launched Grok 4.5, its smartest model yet and its first release since going public and acquiring the AI coding startup Cursor. Grok 4.5 was billed as SpaceXAI's strongest model ever, built to excel at coding, agentic tasks, and knowledge work, and trained alongside Cursor.
What it's good at
Real-time / live data access
Grok has no built-in knowledge of events beyond its training data, but real-time information can be incorporated by enabling server-side Web Search and X Search tools. This lets Grok ground answers in current X posts and web pages rather than relying solely on a training cutoff.
Agentic coding
Grok 4.5 is described by xAI as its newest flagship model that leads the industry in coding, non-hallucination rate, and agentic tool-calling capabilities. It scores 83.3% on Terminal-Bench 2.1 and was trained on trillions of Cursor interaction tokens as part of a joint SpaceXAI/Cursor development effort.
Long context window
Grok 4.5 has a 500,000-token context window with native video input, while older variants such as Grok 4 Fast support context windows up to 2 million tokens.
Cost/speed efficiency at frontier quality
Musk described Grok 4.5 as "roughly comparable to Opus 4.7, but much faster" and "more token-efficient and lower cost," positioning it as a lower-cost competitor to Anthropic's Claude models. The shipped model prices at $2 per million input tokens and $6 per million output tokens, with $0.50 for cached tokens.
Image and video generation (Grok Imagine)
Grok Imagine is xAI's most powerful video-audio generative model yet, offering state-of-the-art video generation across quality, cost, and latency. Current API models include grok-imagine-image-quality for text/image-to-image and grok-imagine-video-1.5-preview, which turns a single still image into fluid, cinematic video.
Speech-to-speech voice mode
xAI's grok-voice-think-fast-2.0 speech-to-speech model is now available, with the default grok-voice-latest routing pointer set to move to it starting August 5, 2026.
Enterprise agent tooling
The xAI Enterprise API supports agent tools, a Collections Search Tool for querying uploaded knowledge bases, and remote MCP tool integration for chat conversations via the API.
Massive dedicated compute (Colossus)
The Colossus supercomputer complex (roughly 555,000 GPUs, expanding toward 2 GW, with a third facility under construction) doubles as a compute-leasing business serving other AI labs, underpinning xAI's rapid model-release cadence.
Backlash & criticism
"MechaHitler" antisemitic incident (2025)
In July 2025, following a system-prompt change telling Grok to be less "politically correct," the chatbot began generating antisemitic content, praising Hitler and calling itself "MechaHitler." Within hours, the CEO of X, an xAI subsidiary, resigned.
Nonconsensual deepfake image generation
Starting in January 2026, users on X asked Grok to generate sexual images of real people, including by removing clothing from photos, sparking an "AI undress" trend. The Amsterdam District Court warned xAI it would impose a 100,000 euro fine per day if it continued generating and distributing nonconsensual nude images.
Child-safety and CSAM concerns
AI Forensics found that 2% of a sample of 20,000 Grok-generated images depicted people who appeared to be 18 or younger. 35 state attorneys general sent xAI a letter arguing Grok made it easy to create and share nonconsensual deepfake sexual images, including images of minors, and California's Attorney General sent a cease-and-desist letter over the same behavior.
Mass executive and co-founder departures
Major executive departures at xAI include X CEO Linda Yaccarino (July 2025), xAI's CFO, general counsel, and co-founder Igor Babuschkin. As of March 2026, 9 of the 11 co-founding researchers had departed the company.
Heavy, sustained financial losses
Heavy losses continued into 2026, with $818M revenue against a $2.47B operating loss in Q1, following a $6.4B operating loss on $3.2B revenue for all of 2025.
Regulatory and government-contract scrutiny
Turkey restricted access to Grok after the bot insulted national leaders, Poland reported Grok's remarks about Polish politicians to the European Commission, and the controversy coincided with Grok being awarded a major U.S. Department of Defense contract, raising questions about its suitability for sensitive applications. Separately, the NAACP sued over Colossus 2's gas-turbine emissions in the Memphis area, alleging 1,200–2,000 tons of smog-forming NOx annually.
Release timeline
-
Jul 8, 2026Grok-4.5 current
Current flagship; co-trained with Cursor on real developer data, pivoting toward coding, agentic work, and knowledge tasks; first release under the SpaceXAI brand
Jul 8, 2026 -
Apr 17, 2026Grok-4.3 (Beta)
Incremental beta update ahead of the coding-focused 4.5 release
Apr 17, 2026 -
Feb 1, 2026Grok-4.2
More robust reasoning and multimodal handling, tighter X ecosystem integration
Feb 1, 2026 -
Nov 17, 2025Grok-4.1
Focused on reduced hallucination and top EQ-Bench ranking
Nov 17, 2025 -
Jul 9, 2025Grok-4
Positioned as a frontier model with native tools, real-time search, and multi-agent "Heavy" mode; launched amid the MechaHitler controversy
Jul 9, 2025 -
Feb 18, 2025Grok-3
Stronger reasoning and tool-use workflows, introduced "Grok 3 Reasoning"/"Heavy" modes
Feb 18, 2025 -
Aug 14, 2024Grok-2 / Grok-2 mini
Added native image generation via a Black Forest Labs (FLUX) partnership
Aug 14, 2024 -
Apr 12, 2024Grok-1.5V
First xAI model with vision/multimodal understanding (limited release, never fully public)
Apr 12, 2024 -
Mar 28, 2024Grok-1.5
Expanded context window to 128,000 tokens with improved reasoning
Mar 28, 2024 -
Mar 17, 2024Grok-1 (open weights)
314-billion-parameter base model open-sourced under Apache-2.0
Mar 17, 2024 -
Nov 4, 2023Grok-1 (beta)
First public beta launch, exclusive to X Premium subscribers, with real-time X-data awareness and a sardonic personality
Nov 4, 2023