The complete guide to Grok AI

Grok is the AI assistant built by xAI, the company Elon Musk founded in March 2023 and that SpaceX acquired in February 2026. Across its website and product announcements the business now brands itself SpaceXAI. Grok is the AI division of the world’s most valuable private company, not an independent startup.

The flagship is Grok 4.6, released on 12 August 2026, one day after xAI launched Grok Bot. Grok 3, Grok 4 and the whole Grok 4 Fast family were retired from the API on 15 May. Pricing has fallen sharply. Grok and X also picked up open investigations in at least six jurisdictions.

This guide covers Grok as it stands in August 2026: how it works, what it costs, where it is strong and where the risks sit.

What is Grok AI?

Grok is a family of large language models and an assistant that runs on grok.com, in the X app, on iOS and Android, and inside Tesla vehicles. Three things separate it from ChatGPT, Claude and Gemini.

• Live access to X. Grok searches the public X post stream as a first-class tool alongside web search. No other major assistant has this.
• A deliberately opinionated personality. xAI trains Grok to answer directly rather than hedge, and it engages with topics other assistants decline.
• Aggressive pricing. At $2 per million input tokens and $6 per million output tokens, the flagship undercuts Claude Opus 5 and GPT-5.6 by a wide margin.

xAI raised a $20 billion Series E in January 2026 and trains on Colossus, its Memphis datacentre, which it describes as 200,000 H100 GPUs in one interconnected cluster.

The design principles behind Grok

xAI’s published mission is short: “Accelerate human scientific discovery.”

The phrase most associated with Grok, maximally truth-seeking, does not appear in that mission. It appears in the system prompts xAI publishes on GitHub, inside a policy block the company gives highest precedence: “You are a helpful, maximally truth-seeking assistant to users around the world and of all ages.” The same block tells the model not to deceive users and to correct false premises.

It is worth being precise about what xAI publishes, because this is widely misreported.

• There is no xAI Model Spec. xAI has published nothing equivalent to OpenAI’s Model Spec or Anthropic’s Constitution. What it does publish is a safety framework, per-release model cards, and the consumer system prompts.
• Grok does not use Constitutional AI. That is Anthropic’s technique, and xAI has never claimed it. The Grok 4.5 model card describes “targeted midtraining and post-training with supervised finetuning and reinforcement learning on human and synthetic reward signals.”
• The safety framework has become less specific, not more. The August 2025 Risk Management Framework set numeric criteria, such as answering fewer than 1 in 20 restricted biological and chemical queries, benchmarked against WMDP, Cybench and MASK. The xAI Frontier AI Framework that replaced it on 30 June 2026 drops those thresholds for qualitative safeguards against NIST AI RMF and ISO/IEC 42001.
• The published prompts lag the product. The newest files in the public repository are Grok 4.1-era. Nothing has been published for 4.5 or 4.6.

In practice, “truth-seeking” is a training objective and a prompt instruction, not an audited guarantee. Treat it as a description of intent.

Grok’s models in August 2026

Model	Context	Input / output per 1M tokens	Best for
Grok 4.6 (flagship)	500K	$2.00 / $6.00	Long-running agents, coding, knowledge work
Grok 4.5	500K	$2.00 / $6.00	Previous flagship, still available
Grok 4.3	1M	$1.25 / $2.50	High-volume work, very long documents
Grok 4.20 Multi-Agent	1M	$1.25 / $2.50	Hard problems worth parallel reasoning
Grok Build 0.1	256K	$1.00 / $2.00	Agentic coding in a terminal or IDE

xAI calls Grok 4.6 “SpaceXAI’s frontier model for coding, agentic tasks, and knowledge work.” It is a post-training upgrade rather than a bigger base model: same 500K context and same 1 February 2026 cutoff, but regenerated fine-tuning trajectories and reinforcement learning in agentic environments such as kernel optimisation, web development and CAD. xAI reports more self-testing on long task chains. It launched on the API, in Grok Build and in Cursor first, and is not yet named on the consumer product page.

The practical change is a new `xhigh` reasoning effort, on Grok 4.6 and later. On Grok 4.5 an xhigh request is silently treated as high. Cached input costs more on 4.6 than on 4.5, $0.50 against $0.30 per million, and long-context pricing still doubles everything above 200,000 tokens.

Note the inversion: the flagship has a smaller context window than the cheaper models. If you need a million tokens, use Grok 4.3 or 4.20.

Retired on 15 May 2026: Grok 3, Grok 4 (grok-4-0709), Grok 4 Fast, Grok 4.1 Fast, Grok Code Fast and Imagine Image Pro. Calls to those IDs bill at Grok 4.3 rates. Grok 5 has not been released. Musk has said Grok 4.7 follows, with no confirmed date.

Where Grok gets its information

The distinction between training and retrieval is the whole answer.

Grok has a fixed knowledge cutoff. xAI’s documentation puts it at 1 February 2026 for both Grok 4.5 and 4.6, while the 4.5 model card says January 2026. Because 4.6 is a post-training upgrade, the cutoff did not move with the new release.

The published system prompt says the opposite. It tells Grok: “Your knowledge is continuously updated – no strict knowledge cutoff.” That is a contradiction between xAI’s own artefacts, and it explains why Grok sometimes insists it has no cutoff.

Anything Grok knows about events after early 2026 comes from retrieval, not the model. Retrieval runs through server-side tools billed separately in the API: web search, X search, code execution and file attachments.

On training data, the best source is the public summary xAI published on 8 July 2026 under EU AI Act transparency rules. It describes a corpus of “publicly available information, including, for example, certain scientific text, legal and official documents, social media posts, and source code,” spanning text, images, audio and video. The web crawler ran from January 2024 to June 2026. xAI says it respects opt-out signals and supplements the corpus with licensed and synthetic data.

What Grok AI can do

Capability	Detail
Real-time search	Live web and X search with citations, plus DeepSearch for deeper research
Multimodal input	Images, diagrams, PDFs and files; vision across all current models
Image and video	Grok Imagine Image 2.0 (August 2026) with multi-reference editing on up to five inputs; Imagine Video 1.5 at native 1080p
Voice	26 voices across 25+ languages, sub-second latency, custom voice cloning via API
Languages	xAI’s product page lists 30+ in chat
Memory	Recalls preferences and past conversations across sessions
Projects and canvas	Group related chats and files into a workspace; canvas for long-form writing
Automations	Describe a job once and Grok runs it on a schedule or an email trigger, then reports back
Build Mode	Generate and publish working sites, apps and games from chat, on web and mobile
Grok Bot	Always-on agents with their own cloud computer that sign into your tools and work unattended. Beta since 11 August 2026, macOS and iOS, SuperGrok Heavy and Cursor Ultra only
Agentic coding	Grok Build, open-sourced July 2026, with MCP support, subagents, hooks and plan mode
Document workflows	Native integrations for Excel, Word, PowerPoint, Outlook and Google Workspace
Enterprise distribution	Amazon Bedrock, Databricks, Azure AI Foundry, GitHub Copilot, Cursor, OpenRouter

For developers, the API adds a Files endpoint (48 MB per file) and Collections, an embedding-indexed retrieval layer with semantic search, configurable chunking and metadata filtering.

How multi-agent reasoning works

Grok’s multi-agent mode is often described as “multi-threaded personalities.” That is misleading. Per xAI’s documentation it is parallel test-time compute with a leader-synthesiser topology.

Multiple instances of the same model reason over your query at once, each able to call tools independently. A designated leader agent synthesises the result into one answer. You control the width: agent_count=4 maps roughly to low or medium reasoning effort, agent_count=16 to high.

Only the leader’s tool calls and final response come back. Sub-agent state is discarded unless you request encrypted content, and you pay for every token every agent consumes. This is not mixture-of-experts routing and not a persistent swarm. It is one model, run wide, with a single instance aggregating. It runs only on grok-4.20-multi-agent: neither Grok 4.5 nor 4.6 supports it.

Plans, pricing and API costs

Plan	Price	What you get
Free	$0	Real-time web and X search, voice mode, connectors. Heavily rate-limited
SuperGrok Lite	$10 / month	Entry tier, longer sessions, basic Imagine
SuperGrok	$30 / month	Flagship model, higher rate limits, image and video generation
SuperGrok Plus	$100 / month	1080p video, substantially higher usage across chat, Imagine, voice and Build, priority access at peak, early access to new features
SuperGrok Heavy	$300 / month	Multi-agent, priority queue, highest limits
Business / Enterprise	Contact sales	Team management, SOC 2, custom terms

Only the Free, SuperGrok and SuperGrok Plus prices appear on xAI’s own pricing page, so confirm Lite and Heavy before budgeting. X Premium is separate: Premium at $8 and Premium+ at $40 per month raise Grok limits inside X, but do not replace a SuperGrok subscription.

On the API side, token prices are only part of the bill. Tools are charged per invocation: web search, X search and code execution at $5 per 1,000 calls, file attachments at $10 per 1,000, and Collections search at $2.50 per 1,000. A batch API gives a 20% discount on Grok 4.3 and the 4.20 family, but not on 4.5 or 4.6. Priority processing costs double.

Getting started with the Grok API

Create an account at docs.x.ai, activate billing and generate a key. The API is OpenAI-compatible, so most client code works with a changed base URL and model name.

import requests
 
response = requests.post(
    "https://api.x.ai/v1/chat/completions",
    headers={"Authorization": "Bearer YOUR_API_KEY"},
    json={
        "model": "grok-4.6",
        "messages": [{"role": "user", "content": "Summarise today's EU AI news."}],
        "reasoning_effort": "medium",
    },
)
print(response.json())

Server-side tools such as web_search, x_search and code_interpreter run through the Responses API at /v1/responses, not chat completions.

Two settings do most of the cost control. Set reasoning_effort to low or medium for routine tasks, since high is the default and burns far more tokens. And keep prompts under 200,000 tokens, because crossing that line doubles the rate.

What Grok AI is good for

Anything where recency decides the answer. Live X and web search make Grok the strongest option for breaking news, market sentiment, competitor monitoring and event tracking. Still its clearest advantage.

Long-running agents and coding. What Grok 4.6 was built for. Artificial Analysis scores it 61 on its Intelligence Index, up from 56 for Grok 4.5 and level with GPT-5.6 Sol. It leads the field on GDPval-AA v2, a knowledge-work test, at 1,753 Elo against 1,526 for its predecessor. On coding it improved sharply without taking the lead: DeepSWE v1.1 rose from 54% to 65.9%, against 73% for GPT-5.6 Sol Max. At $2 and $6 per million tokens, cost per task is a fraction of the competition’s.

Social listening and X analytics. Native access to the X stream makes brand monitoring and trend analysis straightforward in a way rivals cannot match without buying data.

High-volume document work. With 1M context on Grok 4.3 and Collections for retrieval, long contract sets, research corpora and codebases are tractable at low cost. Grok also declines less often than its competitors, which some research teams value and which is exactly why it needs a firmer usage policy.

The downsides: limitations, risk and compliance

Hallucination went up in the last generation. On Artificial Analysis’ AA-Omniscience Index, Grok 4.5’s hallucination rate rose from 25% to 54% against Grok 4.3, while accuracy improved from 35% to 52%. xAI markets Grok 4.6 as having “minimal hallucinations,” but no independent measurement has been published yet. Until one exists, assume the pattern holds.

Real-time access is a double-edged tool. Grounding answers in the live X stream means grounding them in unverified claims. Grok repeats rumours in the same assured tone it uses for established fact.

Regulatory exposure is significant. From December 2025, Grok’s image editing was used at scale to generate non-consensual sexualised imagery of real people, including minors. Note who each action targets, because it matters for procurement.

• European Commission, Digital Services Act proceedings against X, 26 January 2026.
• Irish Data Protection Commission, GDPR inquiry into X Internet Unlimited Company covering Articles 5, 6, 25 and 35, 17 February 2026.
• Ofcom, Online Safety Act investigation into X, 12 January 2026.
• California Attorney General, investigation naming xAI and Grok directly, 14 January 2026.
• Indonesia and Malaysia, temporary access blocks in January 2026.
• A class action against SpaceXAI and Stability AI, amended in July 2026 to cover five plaintiffs.

Most concern the X platform rather than the Grok API, so they do not automatically transfer to an API deployment. They do show how the vendor handles safety failures at scale.

Separately, xAI signed only the Safety and Security chapter of the EU’s GPAI Code of Practice, declining Transparency and Copyright as “profoundly detrimental to innovation.” GPAI enforcement powers under the EU AI Act became active on 2 August 2026.

Data residency is thin. xAI offers a regional endpoint at eu-west-1.api.x.ai, but Grok 4.6’s documented regions are us-east-1 and us-west-2, and data residency at rest requires a conversation with sales.

Governance tooling lags. Audit logging, retention controls and compliance features are thinner than at Microsoft, Google or Anthropic. Check this against your requirements before piloting.

Brand and continuity risk. Grok’s history includes the July 2025 “MechaHitler” episode and several later incidents, and every co-founder Musk originally recruited has left the company. Neither says anything about model quality, but both weigh on a procurement decision.

Grok compared to the alternatives

	Grok 4.6	Claude Opus 5	GPT-5.6	Gemini 3.1 Pro
Input / output per 1M	$2 / $6	$5 / $25	$5 / $30	$2 / $12
Context window	500K	1M	~1M	1M
Live social data	Native X access	No	No	No
Content filtering	Light	Extensive	Extensive	Extensive
Multi-agent mode	Yes, separate model	No	No	No
Enterprise governance	Developing	Mature	Mature	Mature

Prices are for prompts under roughly 200,000 tokens. Grok and Gemini both charge double above that, so compare at the context length you actually use.

Grok is the cheapest frontier model here by a clear margin, and the only one with native access to a live social graph. It also has the thinnest governance story and the most active regulatory file.

Roadmap and what comes next

xAI communicates its roadmap through releases rather than announcements, and it ships fast. Grok 4.5 landed in mid-July, Grok Bot on 11 August and Grok 4.6 the day after. Musk says Grok 4.7 follows, with no date. Grok 5 is reported to be training on Colossus 2, also with no date.

The direction is clearer in the products than the model numbers. In one quarter xAI shipped Build Mode, Automations, Imagine Image 2.0, a voice agent builder, integrations across Excel, Word, PowerPoint, Outlook and Google Workspace, and now agents that run unattended on their own machines. The company is moving from chatbot toward software that executes work, and pricing aggressively to win distribution.

Taking the next steps with Grok AI

Grok in 2026 is a serious option with a specific shape. Fast, unusually cheap, built for long-running agents, and the only assistant with a live view of X. It also carries a hallucination problem in the previous generation and less mature governance tooling than its competitors.

For most organisations the answer is not “Grok or something else” but a multi-model setup where Grok handles what it is best at.

A practical way in:

1. Start with one use case where recency decides the answer, such as competitor monitoring, not a general rollout.
2. Put a verification step between Grok’s output and any decision, especially anything drawn from live X data.
3. Run a data protection assessment before processing personal data, and confirm where requests are processed rather than assuming an EU region.
4. Compare cost per completed task, not headline token prices.
5. Set reasoning effort and context length deliberately. They drive most of the bill.

If you are weighing up which models belong in your stack, our AI consulting service helps you make that call against your own requirements rather than benchmark scores. If you already know and need them safely in production, our AI strategy and development teams take it from there.

Frequently asked questions (FAQ)