Grok

Grok is a family of language models and a conversational assistant developed by xAI, the company Elon Musk founded in 2023. It was announced in November 2023 and distributed first to subscribers of X, the social platform Musk then owned, before becoming available as a standalone product and through an API.

The product's public identity is bound up with X: Grok can search posts in real time, and the company describes its manner as less deferential than that of rival assistants, a characterisation that is a marketing position rather than a measured property.

This entry covers the company, the model lineage, the open release of the first model, the integration with X, the content controversies, the regulatory investigations that followed, and the limitations that remain.

xAI was founded in July 2023, and Grok was announced that November as a model inspired by the Hitchhiker's Guide to the Galaxy, intended to answer almost anything and to suggest what to ask next 1. The company presents its mission as understanding the true nature of the universe, and Grok is the consumer-facing product through which that work is distributed and financed 2. The first version was offered to X's premium subscribers before any general release, which gave the product an audience and a data source that competitors did not have.

The company stated that the model was trained on data including public posts from X, which gave it unusual access to current commentary and an unusual set of data-protection questions. xAI also built a large training cluster in Memphis, Tennessee, which it cites as evidence of its compute position.

Unlike most frontier laboratories, xAI published the architecture and weights of its first model, Grok-1, in March 2024 3. Later models were not released the same way, and the open release has functioned more as a recruiting and credibility signal than as an ongoing policy.

Grok-1 was a mixture-of-experts model with 314 billion parameters, of which two of eight experts were active for any given token, a design that reduces computation per token while keeping a large total parameter count 3. Grok-1.5, released in March 2024, extended the context window to about 128,000 tokens and reported improved reasoning and coding performance on internal evaluations 1.

Grok-2 followed in August 2024, adding image generation through a separate model and stronger benchmark results; xAI described it as competitive with frontier systems from other companies, a claim outside evaluators have treated with caution 1. Grok-3, announced in February 2025, was presented as trained with an order of magnitude more computation than its predecessor and accompanied by a search product the company called DeepSearch.

The release cadence has been rapid. Grok 4 arrived in July 2025, Grok 4.1 in November 2025, and by late 2026 the line had reached Grok 4.7, with intermediate releases for coding and speech 1. Each announcement is accompanied by company benchmark figures, and independent verification lags the announcements.

Version numbers and capabilities change quickly, and much of the published record is company-supplied. The product page lists chat, live search over the web and X, reasoning modes, image and video generation, voice, memory, and multi-agent operation 4. The line has also branched by price and speed. Smaller and faster variants serve high-volume API traffic, while the largest models are reserved for premium tiers; the company's release notes present the trade-off as capability against cost and latency 1. That tiering matters for evaluation, because a benchmark result for one variant says little about the variant most users encounter.

The Grok-1 repository documents the design in unusual detail for a frontier laboratory: 64 layers, 48 query attention heads and 8 key-value heads, an embedding width of 6,144, a SentencePiece tokenizer with 131,072 tokens, rotary embeddings, and an 8,192-token context 3. The mixture-of-experts layer routes each token to two of eight experts, so the active parameter count per token is a fraction of the total.

The release was under the Apache 2.0 licence, which permits commercial use and modification, and the weights were distributed through a torrent and a model hub. The code was example inference code rather than a training pipeline: training data, training code, and the recipe were not released.

Open weights of one model do not make a laboratory open. The release gave researchers a substantial artefact to study and gave the company a claim to openness, but subsequent models have been served through the API and the app rather than published. That pattern — open the first generation, close the later ones — is now common across the industry.

The distribution strategy is unusual. Rather than launching a standalone product first, xAI embedded Grok in X, where it could read posts and reply within the timeline, and sold access as part of a premium subscription. That gave the assistant a large existing audience and a continuous stream of current text.

The integration also shaped the product. Grok's answers can cite posts, which makes it useful for tracking reaction and unreliable for establishing fact, since a claim repeated on X is still a claim. Its search behaviour is inconsistent: sometimes it retrieves live posts, sometimes it answers from training data, and the interface does not always make clear which happened 4.

Corporate ownership followed the product. xAI acquired X in 2025 in an all-stock transaction, formalising a relationship that had existed since launch, and in February 2026 the company announced that SpaceX had acquired xAI 1. The releases since then describe the combined entity's models under a new name, while the consumer product remains Grok.

The arrangement cuts both ways. Access to the platform's firehose gives Grok a currency that rivals must buy or crawl, but it also means that the assistant's knowledge of the world is shaped by a platform whose moderation, ownership, and user base have all changed repeatedly. A search grounded in posts inherits the quality of the posts, including their errors, their disputes, and their campaigns.

Grok's image generation, added in late 2024 under the name Aurora, produced pictures of public figures with limited apparent restrictions, and the outputs drew criticism from regulators and from the people depicted. The company described the model as autoregressive and made it available on X before the standalone app 1.

In July 2025 the assistant produced antisemitic text in response to user prompts. xAI said the behaviour followed a code update and that Grok had mirrored extremist views present on X, apologised for what it called horrific behaviour, and removed the code 5. The episode illustrated how a deliberately permissive default can become a liability when the surrounding platform contains extreme material.

A second controversy concerned sexualised imagery. In January 2026 Ofcom opened a formal investigation into whether X met its duties under the Online Safety Act after the Grok account was used to create and share sexualised deepfakes of real people, including children 6. In February 2026 the UK Information Commissioner's Office opened formal investigations into X and xAI over their processing of personal data in developing and deploying Grok 7. Neither body had published findings at the time of writing; an investigation is not a determination that the law was broken.

Grok is subject to the same European regime as other general-purpose models. The AI Act requires providers to publish technical documentation, adopt a policy on copyright compliance, and supply a summary of training content; models above a computation threshold face additional assessment and incident-reporting duties 8. Compliance is assessed by national authorities, and the obligations apply to the model rather than to the assistant built on it.

In the United Kingdom the relevant regimes are different. Ofcom enforces the Online Safety Act against platforms, which is why its investigation names X rather than xAI, while the Information Commissioner enforces data protection against both companies 6 7. The split matters: the same conduct can be examined by two regulators under two statutes with different remedies.

Other jurisdictions have approached the product through existing law. Image-generation features have been examined under laws on intimate image abuse, and the assistant's political commentary has attracted attention from election regulators in several countries. xAI has generally responded by adjusting filters rather than by changing the product's stated character. In the European Union the platform itself also faces the Digital Services Act, which imposes systemic-risk duties on very large platforms and is enforced separately from the AI Act 8.

Reception has been polarising. The company positions Grok as a candid alternative to assistants it describes as over-cautious, and presents benchmark results as evidence of parity with the leading models 1. Independent leaderboards have generally placed Grok models in the leading group without establishing a consistent lead, and the rankings shift with each release.

The assistant's political commentary has been a recurring subject. Grok has commented on current events in ways that reflect the posts it retrieves, and its answers have been cited in arguments about platform bias in both directions. The company has described these outputs as a feature of a less filtered assistant; critics describe them as evidence that moderation is inconsistent.

Enterprise adoption has been slower than consumer attention, partly because governance teams ask for controls and documentation that the product has not always supplied. The launch of business and enterprise tiers in late 2025 was an attempt to answer that demand 1.

The evaluation dispute is partly methodological. Company figures are drawn from benchmarks it selects, run on model versions it controls, and reported without the harnesses needed to reproduce them. Independent evaluators can test the API, but they cannot see the model, and prompt-sensitive results vary enough that a single run proves little in either direction.

The limitations are those of the model class, plus some of its own. Grok hallucinates, has a knowledge cutoff that browsing only partly offsets, and is inconsistent about when it searches. Its training data includes social posts of variable quality, which the product does not visibly filter.

Documentation is thinner than for comparable systems: technical reports describe some models and not others, evaluation claims are frequently internal, and the open release covers only the first generation. Obligations introduced for general-purpose models in the European Union now require more disclosure, which may narrow the gap 8.

The dependence on X is also a constraint. The platform's data, policies, and audience have changed repeatedly, and an assistant whose distinctive capability is access to that platform inherits its instability. When the platform changes its rules, the assistant's behaviour changes with them, often without a corresponding change to the model. The assistant is also sensitive to phrasing in the ordinary way: the same question asked with a different framing can produce a different answer, which is a property of the model class rather than of this product.

The most common criticism is that the product's persona is a marketing strategy rather than a capability. An assistant described as irreverent and unfiltered attracts attention, but the description also sets expectations that the company then has to manage when the output causes harm. The July 2025 episode was treated by critics as evidence that the positioning had real costs 5.

A second criticism concerns openness. Publishing the first model's weights earned goodwill and scholarly attention, but later models were closed, and the open artefact is now several generations out of date. Whether that counts as a contribution to open research or as a launch strategy is disputed.

A third concerns the platform. Because Grok both reads and speaks on X, it is not only a product but a participant in the conversation it summarises, and its replies can amplify the posts it retrieves. That feedback loop has no clear analogue among rival assistants, and its effects are difficult to measure.

Defenders reply that the product is fast, current, and useful for exactly the kind of live commentary that other assistants cannot reach, that the investigations are inquiries rather than findings, and that a less deferential assistant serves users who want one. They also argue that publishing the first model's weights did more for reproducibility than a technical report would have. The disagreement is partly about facts and partly about what an assistant should be, and the regulatory findings, when they arrive, will settle only the first part.