The short version: what Grok is in September 2026
Grok is the model family built by xAI. SpaceX acquired xAI on February 2, 2026, and the combined company now operates as SpaceXAI. Product names did not change: Grok still runs on grok.com, inside X, in the mobile apps, and through the xAI API.
The current flagship is Grok 4.6, released August 12, 2026, with a 500,000-token context window and configurable reasoning effort. SpaceXAI positions it for coding, long-running agents and knowledge work. It is also generally available on Amazon Bedrock, Microsoft Foundry and Google's enterprise model catalog, which matters if your procurement runs through one of those clouds.
Businesses can buy three things here, with three very different risk profiles: seats (Grok Business and Grok Enterprise), Grok Bot (agents that sign into your tools and work unattended), and the API. A free consumer tier exists for trying it, but free is not the tier you deploy on.
The honest summary: Grok is a credible frontier option with one capability nobody else sells, aggressive per-token pricing, and a vendor in the middle of a large corporate reorganisation. It is a sound choice for some workloads and a poor one for others.
Who makes Grok now, and why that matters commercially
The xAI acquisition closed in early February 2026 and was reported as the largest merger on record; the operating entity is SpaceXAI LLC. In August 2026 SpaceXAI also completed its acquisition of Cursor, folding that team into work on Grok Build, Grok Bot, the Grok API and Cursor itself. That is why Grok Bot access is bundled into Cursor subscription tiers as well as SuperGrok tiers.
For a buyer this is a procurement fact, not a verdict on the technology. Two acquisitions in seven months and a reported IPO process mean roadmap and pricing are less settled than at a vendor shipping the same enterprise contract for five years. Ask about contract continuity, notice periods on price changes, and what survives another restructuring.
Where Grok is genuinely competitive
Three areas hold up: long-horizon agentic work, real-time search across the web and X, and price at the frontier tier.
Long-running agents and coding
Grok 4.6 was tuned for work that runs over many steps rather than single answers: research, analysis, building a working first version of an application, and checking its own output partway through. SpaceXAI reports it matching GPT-5.6 Sol on the Artificial Analysis Intelligence Index. That is a vendor-reported figure, so treat it as a reason to run your own evaluation, not as a result.
The test that matters is cheap: give two or three models the same real task, with your real messy inputs, and score the outputs blind. Grok Build, SpaceXAI's terminal coding agent, defaults to Grok 4.6 and supports MCP, so the route to production is short if your developers already work in Cursor.
Real-time search across X is the differentiator
The xAI API exposes server-side tools the model calls mid-response, including web search and a dedicated X search. Because SpaceXAI and X are under common ownership, this is first-party access to the live platform rather than a scraped or secondhand feed.
Useful cases: brand and competitor monitoring, competitive analysis and SWOT input that reflects this week rather than last quarter, catching an incident while it is still a complaint thread, and routing what you find into support tickets. The limit is that X is a biased sample of the public, skewed by who posts and what gets amplified. It is a fast signal, not market research.
Context size and price position
Grok 4.6 carries a 500,000-token context window. The older grok-4.3 and grok-4.20 families carry a one-million-token window at roughly half the price, which makes them a sensible destination for high-volume, lower-difficulty work in a tiered setup.
Pricing is also tiered by prompt size, and the step is steep: rates roughly double once a prompt crosses 200,000 tokens. Context discipline is a line item, not a style preference. A retrieval layer that sends 8,000 relevant tokens instead of 250,000 is cheaper twice over.
Grok Bot: what an always-on AI teammate actually does
Grok Bot launched in beta on August 11, 2026 and is both the most interesting and the most dangerous thing SpaceXAI currently sells to businesses. It is the clearest live example of the shift from assistants that answer to systems that act, a pattern our complete guide to agentic AI covers independent of any one vendor.
How it works
You create named Bots that run on a persistent cloud computer with a browser, a filesystem and a terminal, and keep working when you close your laptop. A Bot signs into web applications the way a person does, so it can operate tools with no API at all. Where a site demands a CAPTCHA or two-factor prompt, the documentation says the Bot hands control back to you rather than trying to get around it.
Three building blocks sit on top. Connectors are installed from Settings to Plugins and wire a Bot into supported systems. Skills are reusable written instructions for a task, including decision rules, expected output and safety boundaries. Routines decide when a workflow runs, on a schedule or a supported event, capped at 50 per Bot with the 20 most recent run records retained.
Access is bundled into subscription tiers rather than sold standalone: SuperGrok Plus and Heavy, Cursor Pro+ and Ultra, and Cursor Teams Standard and Premium, with enterprise deployment rolling out separately. Those tiers and prices have moved repeatedly since launch, so check the current plan page rather than any figure in an article.
One detail is widely misreported. Bots do not each get a private machine: SpaceXAI's documentation states that all Bots on an account share one cloud computer with shared files, browser sessions and logins, and warns plainly not to use separate Bots as a security boundary. Isolating two trust domains requires separate accounts.
What unattended access to your tools actually implies
Granting an agent a login is not the same as granting a person one. Anything a signed-in human could do in that tool, the agent can do, with no hesitation before hitting delete and no colleague glancing at the screen.
The agent also reads untrusted content all day: inbound email, web pages, support tickets, PDFs from strangers. Text inside that content can be crafted to redirect its behaviour. Prompt injection is unsolved across the industry, and there have been publicly reported incidents where injected instructions produced unintended actions through Grok integrations. The risk belongs to every computer-use agent from every vendor, but it is live, not theoretical.
Then the audit trail. Actions taken through a shared login are attributed to whoever owns it. If an agent emails a customer the wrong invoice, you want the logs to settle that beforehand.
The guardrails to demand before you switch it on
Use dedicated service accounts, never a staff member's login, scoped to the minimum permissions the job needs. A bookkeeping Bot does not need refund rights. Separate trust domains with separate accounts, not separate Bots.
Make approval rules mandatory for the irreversible categories: sending external messages, publishing, deleting data, purchasing, and anything that changes production. SpaceXAI's documentation supports and recommends exactly this. Treat the recommendation as a requirement, and keep payment rails, wallets and credential managers out of reach entirely.
Run read-only for the first few weeks, with the agent drafting and a human sending. Log every run and sample-audit them. Test each skill manually before attaching a routine, and define what it does when its source data is stale, because an agent running on schedule against nothing will confidently produce something. Write the rollback first: how to revoke sessions, rotate credentials and stop every routine.
Grok Business vs Grok Enterprise: the plans and what they include
SpaceXAI sells organisational access in two tiers, announced at the end of December 2025. Grok Business is $30 per user per month, self-serve, adding team management, consolidated billing and user analytics on top of the same models individuals use. It suits small teams that want dedicated workspaces for personal and team use rather than everyone expensing a personal SuperGrok subscription.
Grok Enterprise is custom pricing through xAI sales and adds the controls larger organisations ask for: custom SSO, directory sync via SCIM, role-based access control, and advanced audit and security controls. Enterprise Vault is an optional add-on on that tier providing a dedicated data plane, application-level encryption and customer managed encryption keys. If your requirement is technical isolation from shared consumer infrastructure rather than a written promise, that add-on is the line you want.
Both tiers integrate with tools like Google Drive through Apps, with permission-aware access so the assistant only sees what the requesting user can already see, and cited sources. Projects run agentic search over large document stores via the Collections API.
Roll it out account-first: register the workspace on a company domain with a named internal owner, so licences, conversations and audit logs belong to the business, not to individuals. Then write the use policy before the rollout email, covering what data is allowed in prompts and which surface staff should use.
What the xAI API costs, and what actually drives the bill
The API is usage-based per token, billed separately from seats, with different input and output rates and a cheaper rate for cached input. As of September 2026 the published rate card lists grok-4.6 at $2.00 per million input tokens, $0.50 cached input and $6.00 output for prompts under 200,000 tokens, doubling above that threshold; grok-4.3 and the grok-4.20 family sit at $1.25 input and $2.50 output. Server-side tools are billed per 1,000 invocations: $5 each for web search, X search and code execution, $2.50 for collections search, $10 for file attachments. Verify at docs.x.ai before budgeting, because these numbers have moved repeatedly this year.
The mechanism is what stays true. Three things drive your bill and none is the headline rate: tokens sent per call, whether you reuse cached context, and whether every task goes to the most expensive model. Routing extraction and classification to a cheaper model usually saves more than switching vendors.
For a typical small-business automation, model spend is not the expensive part; building the system is. A custom build at AutomateNexus starts at $7,500 with a typical 30-day timeline, or four to eight weeks for a larger MVP, and ongoing model cost usually lands in the $30 to $150 per month range paid directly to the provider.
Is Grok safe to use with business data? The governance questions
The answer differs completely by surface, and people conflate the surfaces constantly.
For the xAI API, SpaceXAI's security documentation states it does not train on API inputs or outputs without explicit permission, and that requests and responses are stored encrypted for 30 days for abuse auditing, then deleted. Zero Data Retention is an enterprise option that stops prompts and outputs being written to disk at all, with a trade-off the documentation is upfront about: it disables the stateful Responses API, the Files API, Collections and the Batch API.
Consumer surfaces are a different regime. X's January 2026 terms treat Grok prompts and outputs as user content usable for training and fine-tuning. Opt-out controls exist but live in separate settings screens for Grok on X, the mobile app and grok.com, and switching them off does not reach back to data already collected. So "we use Grok" means very different things depending on where your team types.
SpaceXAI claims SOC 2 Type 2, GDPR and CCPA compliance, states that business and enterprise customer data is not used for training, and says HIPAA is not standard, with a Business Associate Agreement available on request. Verify each claim through the Trust Center, and get the commitments into the contract rather than an FAQ.
The questions to put in writing, for any vendor: does the contract itself say you will not train on our data? What is the retention window, and can it be zero? Which subprocessors touch the data? Who at the vendor can read agent session logs, given those logs contain everything the agent saw inside your tools?
Grok vs ChatGPT, Claude and Gemini for business
Compare on fit, not leaderboard position. At the frontier tier the capability gaps are narrow enough in 2026 that ecosystem, data terms and performance on your specific inputs decide more outcomes than intelligence scores do.
ChatGPT remains the safest default when the goal is one tool a non-technical team will actually adopt, on ecosystem, integrations and staff familiarity. Anthropic's Claude is the common pick for document-heavy professional work and coding agents. Gemini wins mostly on gravity: if you already live in Google Workspace, the path of least resistance is worth something. Our side-by-side on Claude versus ChatGPT for business goes deeper on the first two. If data residency or running without any vendor is the constraint, the honest answer is not a hosted API at all, and the guide to the best open-source LLMs for business covers that path.
Grok's advantages are specific: live X data nobody else can sell, frontier pricing at the lower end of the majors, tuning for long-running agent work, and a short route to production if your engineers use Cursor. Its disadvantages are equally specific: a shorter enterprise procurement track record than OpenAI, Anthropic or Google, a smaller third-party integration ecosystem, consumer-surface data defaults that need active management, and a brand some customers hold an opinion about.
Brand perception is a real deployment variable
Stated plainly and then left alone: Grok is a more polarising brand than its competitors for some audiences, and that is a commercial input rather than a technical one. For internal and back-office work it is close to irrelevant. For customer-facing deployments where the vendor is named, it is a branding decision to test with your actual segment. Systems built on the API do not surface the underlying model by default, so for most custom builds it only arises if you disclose it.
Who should consider Grok, and who should not
Consider it if your team already pays for Cursor or SuperGrok and Grok Bot access is effectively bought; if you need live X data and currently get it late or manually; if you run high-volume agent workloads where per-token price compounds; if you have engineering capacity to own an API integration; or if your first target is internal back-office work where a mistake has a contained blast radius.
Skip it, for now, if you handle protected health information without a signed BAA; if your buyers or board require a long, stable vendor track record; if you want a turnkey no-code product with a large marketplace of prebuilt connectors; if you will not run approvals and dedicated service accounts on an unattended agent; or if you sell into a segment where the vendor brand becomes part of the conversation.
For everyone in between: pick one workflow that costs real hours, run it through Grok and two competitors with your own data, and compare quality, total cost and integration effort. Build on the winner in your own account with your own key, so the decision stays reversible. That bring-your-own-key arrangement is the most consequential contractual decision in an AI build, and it is how AutomateNexus works by default: model-agnostic, client holds the provider account, no markup on usage. Standardising how the model reaches your internal tools is what MCP is for.
If you would rather start from the workflow than the model, the free automation audit is a three-minute questionnaire returning an automation health score, what manual work costs you annually, and a ranked list of quick wins. No call, no cost, and deliberately model-agnostic.
Grok for business FAQ
Prices and version numbers below are accurate as of September 2026; re-check them against SpaceXAI's own pages.
Is Grok owned by SpaceX now?
Yes. SpaceX acquired xAI on February 2, 2026 and the combined entity operates as SpaceXAI LLC. In August 2026 it also completed its acquisition of Cursor. Access points did not change: grok.com, X, the mobile apps and the xAI API.
Is Grok free for business use?
There is a free consumer tier, fine for evaluating the model. It is not a business deployment: that means paid seats under Grok Business or Grok Enterprise, or API spend under your own account. Free tiers also carry the loosest data terms.
How much does Grok cost for a business?
Grok Business is $30 per user per month self-serve; Grok Enterprise is custom pricing through xAI sales. API usage is billed separately per token, with grok-4.6 at $2 per million input and $6 per million output tokens below the 200,000-token threshold, plus per-thousand charges for server-side tool calls.
What is the difference between Grok Business and Grok Enterprise?
Business is the self-serve team tier: team management, consolidated billing, user analytics. Enterprise adds custom SSO, SCIM directory sync, role-based access control and advanced audit and security controls, with Enterprise Vault available as an add-on for a dedicated data plane, application-level encryption and customer managed encryption keys. If you need technical isolation rather than a written promise, Enterprise Vault is the line you want.
Does Grok train on my business data?
It depends on the surface. SpaceXAI states it does not train on API inputs or outputs without explicit permission, with 30-day default retention for abuse auditing and a Zero Data Retention option on enterprise agreements, and applies the same no-training position to business and enterprise accounts. Consumer Grok on X and grok.com operates under terms that do allow prompts and outputs to be used for training, with per-surface opt-outs. Read the current terms for the surface you are using.
Can enterprise developers build on Grok through the API?
Yes. The xAI API is REST at api.x.ai/v1 and is compatible with both the OpenAI and Anthropic SDKs, so moving an existing integration across is a base URL change and a new key. It supports server-side tools including web and X search, MCP for reaching internal systems, and multi-agent model variants. Grok 4.6 is also on Amazon Bedrock, Microsoft Foundry and Google's enterprise catalog.
Is ChatGPT or Grok better for business?
For a general-purpose tool a whole non-technical team will use, ChatGPT is still the lower-risk default on ecosystem and familiarity. Grok is the better answer for live X data, cost-sensitive high-volume agent workloads, and teams already inside Cursor. Neither generalises across every workflow, which is why the head-to-head on your own data is worth an afternoon.
Should we switch our whole company to Grok?
Not as a blanket move. Retraining a company on a new chat tool costs more than a modest capability difference repays. Keep what staff already use for general work, and choose the model per system when you build something specific. SDK compatibility makes that comparison a configuration change rather than a project.