Grok Bot Pricing 2026: Real Costs, Limits, Alternatives

A guide to what xAI's Grok products actually cost in 2026, from the free @grok tag on X to the $300/month agent tier, and where the real limits and cheaper alternatives are.

On August 11, 2026, xAI shipped a new product called Grok Bot at a $200-to-$300-a-month entry price, then quietly dropped that floor three times in fifteen days. By August 26 Cursor's own pricing page listed Grok Bot inside its $20/month Pro plan, while xAI's own Grok Bot FAQ still listed a $60-to-$100 floor. Two companies selling the same product, on the same day, could not agree on what it cost.

That is not a one-off pricing glitch. It is the story of Grok pricing in 2026: a chatbot subscription ladder that most people can describe accurately, wrapped around a brand-new autonomous agent product whose access tiers, usage limits, and even its name ("Grok Bot" now means three different things depending on which xAI page you read) have moved every few weeks since launch. Anyone searching for a straight answer on what Grok actually costs runs into stale screenshots, conflicting aggregator tables, and two primary sources that disagree with each other.

This guide untangles it using only what xAI, X, and Cursor publish on their own current pages, cross-checked against reporting from the weeks the prices actually changed. It covers the full consumer ladder (free tier through SuperGrok Heavy), the new Grok Bot agent product and its real limits, the developer API, Grok for teams, the security model xAI itself warns about, and where the credible alternatives sit on cost and capability.

Contents

1. What "Grok Bot" Actually Means: Three Different Products
2. The SuperGrok and X Premium Ladder: What You Pay for Chat, Voice, and Imagine
3. Grok Bot: The New Always-On Agent, and Its Three-Week Price Collapse
4. The Real Limits: Weekly Usage Pools, No Spend Cap, and the Bills That Surprised Early Users
5. Security and Failure Modes: The Shared-Computer Problem
6. The Developer API: Grok 4.6 Token Pricing for Builders
7. Grok for Business and Enterprise: Per-Seat Pricing for Teams
8. Alternatives: AI Chatbot Subscriptions
9. Alternatives: Always-On Agent Platforms
10. How to Choose: A Decision Framework by Use Case

Before the section-by-section breakdown, the table below scores the six always-on agent platforms this guide covers on the criteria that actually separate them: what it costs to unlock the capability, how reliably it finishes real tasks, how transparent the vendor is about usage limits, how many tools and apps it can actually operate, and how contained the security model is when something goes wrong. Every cell carries its own justification, not just a number, and the row order is the ranking.

Weights: Entry Cost 25%, Reliability 25%, Usage Transparency 20%, Platform Reach 15%, Security Model 15%. Formula: (score1 x weight1) + (score2 x weight2) + ... Scores are 0-10, 5 is adequate, 10 is best in class. Rows 3 and 4 tie at 6.9 and are ordered alphabetically.

Grok Bot lands at the bottom of this table, and it is worth sitting with why, since it is also the newest and in raw capability terms the most ambitious entry on the list. Full computer use across any app without an API is a genuine edge over a browser-sandboxed agent or a code-scoped one, which is exactly why it still scores a 9 on Platform Reach. But a product that will not tell you your weekly allowance, offers no spend cap, and shares one set of credentials across every Bot an account runs cannot score well on the two criteria carrying the most weight in this table, Reliability and Usage Transparency together account for 45% of the final number, and Grok Bot is the lowest-scoring platform on both. The rest of this guide walks through exactly why, starting with what "Grok Bot" even refers to.

1. What "Grok Bot" Actually Means: Three Different Products

Anyone typing "Grok Bot pricing" into a search bar in September 2026 is almost certainly looking for one of three unrelated things, and conflating them is the single most common mistake in coverage of this topic. xAI, now operating as a subsidiary of SpaceX under the SpaceXAI brand after a February 2026 acquisition at a roughly $250 billion valuation, kept the consumer-facing Grok name unchanged through that corporate restructuring - Wikipedia. The company name changed. The product names underneath it did not, which is exactly why they are so easy to mix up.

The oldest of the three is the @grok reply bot on X, a feature that has existed since March 2025: tag @grok in a reply and it answers using the context of the post you tagged, including summarizing video, checking claims, or explaining an image - Layer3 Labs. As of March 2026, that specific feature requires an X Premium or Premium+ subscription to trigger; it no longer works for logged-out or Basic-tier accounts. The second is SuperGrok, the standalone consumer chat subscription (grok.com and the Grok mobile app) covering text chat, DeepSearch, Voice mode, and Grok Imagine for images and video. The third, and the one that actually generated the pricing chaos this guide investigates, launched on August 11, 2026: Grok Bot, xAI's persistent, always-on computer-use agent that signs into a user's real software and keeps working after the device is closed - x.ai official announcement. xAI itself has raised billions of dollars across multiple funding rounds since its founding to build toward exactly this kind of agentic product line, a trajectory O-mega covered in detail back when the company was still purely a model lab rather than a SpaceX subsidiary - xAI, Elon Musk's AI company, has raised $6 billion.

The distinction matters because the three products have completely different cost structures, completely different risk profiles, and completely different audiences. A casual user asking Grok to summarize a thread on X pays nothing beyond their existing Premium subscription and faces essentially no security exposure. Someone running Grok Bot to manage a CRM overnight is running unattended software with standing credentials to their real tools, billed against a usage pool that, as later sections cover, xAI does not fully disclose. Mixing the two in a single mental model, which most of the SEO aggregator content on this topic does, produces exactly the kind of "$8/month to $300/month, take your pick" confusion that makes Grok pricing feel arbitrary when it is not.

Why this matters in practice is that pricing questions have different correct answers depending on which product a reader means. "How much does Grok cost" for the reply-bot use case is answered entirely by the X Premium tier table in the next section. The same question for the agent use case requires understanding a beta rollout that changed its access requirements three times in fifteen days, covered in Section 3. Readers who only need the former can skip directly to Section 2 and never touch the more volatile, more expensive, and considerably riskier agent product at all.

The naming collision is not an accident of casual usage either. xAI's own release notes use "Bot" as the product name in the exact same sentence structure that its FAQ uses for the older @grok feature, and neither the release notes nor the launch announcement acknowledge that the term was already load-bearing elsewhere in the product line - xAI Developer Release Notes. That is a genuine documentation gap, not a reader's misunderstanding, and it means a support search, a Google query, or a colleague's offhand mention of "the Grok bot" carries real ambiguity that this guide's disambiguation is meant to resolve before any pricing number gets attached to the wrong product.

It helps to walk through how a typical reader actually arrives at each of the three. Someone scrolling X and seeing a reply thread where another user tagged @grok to fact-check a claim is looking at the first product, and the only decision in front of them is whether an $8/month Premium subscription is worth it for that feature plus the rest of X's paid perks. Someone who has outgrown free-tier Grok's rolling rate limit and wants a dedicated assistant for writing, research, or image generation is shopping the second product, the SuperGrok ladder covered in full in the next section. Someone who has read about an AI agent that can log into a CRM overnight and file a report by morning, and is trying to figure out what that costs, is asking about the third product, and is the reader this guide's middle sections were written for most directly.

2. The SuperGrok and X Premium Ladder: What You Pay for Chat, Voice, and Imagine

The consumer chat product is the part of Grok's pricing that xAI documents most consistently, and it has settled into seven distinct price points spread across two parallel purchase paths: buying Grok directly through xAI (SuperGrok), or buying it bundled into an X subscription. Both paths exist because X and xAI, while now sibling entities under the same SpaceXAI umbrella, still sell access separately, and the two ladders do not map onto each other cleanly. A Premium+ subscriber and a base SuperGrok subscriber both pay roughly the same amount but get different bundles of value: one gets X's ad-free timeline and creator tools plus Grok, the other gets nothing from X but a materially larger Grok usage allowance.

Free access exists but is deliberately thin. xAI runs free users on a lighter model (commonly referred to as Grok 4 Mini in current documentation) with no image generation and a rolling-window rate limit rather than a fixed daily cap, meaning the reset clock starts from your first message in a window rather than at midnight - DataStudios. That structure is easy to misjudge from the outside: a guide describing "10 free messages a day" is describing a system that no longer exists as of mid-2026, since the reset is now tied to a rolling window measured in hours, not a calendar day.

Two structural facts explain why this ladder looks the way it does. First, X Premium's Basic tier ($3/month, adds post editing and longer posts) deliberately excludes Grok entirely, which means the true floor for any Grok chat access at all is $8/month, not $3 - XreplyAI. Second, X restructured its own Pro dashboard access in March 2026, moving what used to be available to standard Premium subscribers behind the Premium+ paywall without advance notice, a five-times price jump for anyone who relied on that specific feature - MacRumors. That change is a useful signal for the rest of this guide: xAI and X have both shown a pattern of moving feature-to-tier mappings after launch, which is exactly what happened again with Grok Bot five months later.

A single structural change reshaped how every one of these tiers is measured. In June 2026, xAI replaced the older system of separate daily caps per feature (one limit for chat, another for image generation, another for voice) with a single shared weekly usage pool spent across Chat, Imagine, Voice, and Build - CloudZero. Under the old system, running out of image generations for the day did not touch your chat allowance. Under the current system, a heavy afternoon of Grok Imagine video generation can leave a SuperGrok subscriber unable to send a normal text message until the weekly reset, because video consumes disproportionately more of the shared pool than a short chat message does. xAI has not published the size of that pool for any tier, so the practical answer to "how many messages do I get on SuperGrok" is a percentage meter in account settings, not a number you can plan around in advance - Christopher Alarcon.

Grok Imagine itself deserves a specific note because its resource cost inside that shared pool is asymmetric. The feature generates 1K and 2K resolution images and 480p-to-720p video up to 15 seconds long with native synced audio, and it supports text-to-image, image-to-video, and video editing on clips up to roughly 8.7 seconds - PixVerse. On the developer API side (covered fully in Section 6), that same video generation is metered separately at $0.08 per second of output, which gives a rough sense of just how much compute a single video clip represents relative to a text response - UlazAI. That asymmetry is the mechanical reason the shared weekly pool feels unpredictable: a feature most users treat as occasional and lightweight (generate one video) can consume a week's worth of chat headroom in a single request.

A worked example makes the tradeoff concrete. Picture a SuperGrok subscriber who spends a normal week sending roughly 200 short text messages, then on Tuesday afternoon generates three 10-second Grok Imagine clips for a social post. At the API's disclosed per-second video rate, three 10-second clips represent about $2.40 worth of underlying compute, comparable to well over a thousand ordinary chat exchanges at typical token counts. Because the subscription draws from one shared pool rather than separate meters, that single afternoon of video generation can leave noticeably less chat headroom for the rest of the week than the raw message count would suggest, and the only visible warning is a percentage ticking down in account settings rather than a breakdown of which feature spent it.

Two feature names inside the SuperGrok tier and above are worth defining rather than leaving as marketing labels, since both determine what a subscriber is actually paying for. Big Brain mode routes a query through additional reasoning steps before responding, intended for harder analytical or multi-step problems where the default model would otherwise answer too quickly and too shallowly. Expert mode similarly extends response depth and citation density for research-style queries. Neither is metered separately from the shared weekly pool described above; both simply consume it faster than a default chat response would, which is worth knowing before toggling either one on for a task that does not actually need the extra depth.

3. Grok Bot: The New Always-On Agent, and Its Three-Week Price Collapse

Grok Bot is the newest and most consequential thing xAI has shipped under the Grok name in 2026, and it is also the part of this guide's pricing story that changed the fastest. The product itself is a genuine departure from a chatbot: each Grok Bot a user creates runs on its own persistent cloud computer, complete with a real browser, filesystem, and terminal, and it keeps executing tasks around the clock even after the user's own laptop is closed - x.ai official announcement. Its signature capability is "teach-a-task": a user demonstrates a workflow once while the Bot observes, and the system converts that demonstration into a reusable, scheduled routine without writing code or wiring up an API integration - VentureBeat.

The product's technical foundation traces directly back to a corporate deal that closed just weeks before launch. SpaceX finalized a $60 billion all-stock acquisition of Anysphere, the company behind the Cursor code editor, on August 14, 2026, three days after Grok Bot's beta debut - Cryptopolitan. That deal, itself the largest startup acquisition on record at the time, resulted in roughly 391 million newly issued SpaceX Class A shares changing hands and made Cursor a wholly owned unit of the new SpaceXAI division rather than an independent company partnering with xAI at arm's length - Yahoo Finance. That distinction matters for anyone trying to predict where this pricing structure goes next: a partnership can be unwound or renegotiated, but a wholly owned subsidiary's pricing decisions answer to one parent company's incentives, which is a large part of why Grok Bot access moved so freely between the two brands' pricing pages in its first weeks.

That is not a coincidence of timing. Grok Bot's computer environment runs on Cursor's cloud-agent infrastructure, and access to the product is sold through both companies' subscription ladders simultaneously, which is the direct cause of the pricing confusion the rest of this section documents - reworked.co. Cursor's own team pricing adds a third path into the same product beyond the individual tiers already covered: Cursor Teams Standard runs $40 per seat per month with trial-only or on-demand Grok Bot access and no included allowance, while Cursor Teams Premium at $120 per seat adds centralized billing, a shared team marketplace, usage analytics, and SAML or OIDC single sign-on alongside a real included allowance - eesel AI.

At launch on August 11, 2026, Grok Bot was gated behind the most expensive tier on each side of that ladder. Individual access required SuperGrok Heavy at $300/month or Cursor Ultra at $200/month; team access required Cursor Teams Premium at $120 per seat per month - VentureBeat. That positioning read as deliberate: a beta limited to the customers already paying the most, and therefore the most tolerant of an unpolished product and an undisclosed usage meter.

It did not stay there. Reporting on August 21 placed the individual floor at SuperGrok Plus ($100/month) or Cursor Pro+ ($60/month), and xAI's own current Grok Bot FAQ, fetched directly from docs.x.ai for this guide, still states exactly that set of eligible plans: "SuperGrok Plus, SuperGrok Heavy, Cursor Pro+, Cursor Ultra, and Cursor Teams Standard and Premium" - xAI Grok Bot FAQ. By August 26 or 27, coverage of a further expansion appeared, reporting that "all SuperGrok and Cursor Pro subscribers now have access to the Grok Bot," which would put the individual floor at $20/month via Cursor Pro or $30/month via base SuperGrok - KuCoin.

xAI's own launch materials describe the intended use cases in concrete terms rather than abstractions, and the examples are worth repeating because they clarify what a subscriber is actually paying for at each tier. One illustration is a sales outbound Bot that researches target accounts overnight, scores each contact for buying intent, and drafts outreach emails and LinkedIn messages written in that specific seller's own voice, all before the sales team's next workday starts. A second is a demo-readiness Bot that checks a product environment overnight, repairs broken data seeds or stale test accounts, and leaves a checklist waiting before the first customer call of the morning - x.ai official announcement. Both examples share a pattern: the value is not a single answer to a single question, it is unattended, multi-step work completed on a schedule the human never has to babysit, which is precisely the capability none of the SuperGrok chat tiers in Section 2 offer at any price.

That last figure is where the two companies' own current pages stop agreeing with each other, and it is worth stating plainly rather than picking a side. Cursor's own pricing page, fetched directly for this guide, lists "Grok Bot access" as a feature of the base $20/month Pro plan - Cursor. xAI's Grok Bot FAQ, fetched the same day, does not list Cursor Pro or base SuperGrok among eligible plans at all. Both pages are live, both are the vendor's own current documentation, and they describe different products. The most defensible reading, given the reporting trail above, is that xAI did roll access down to the $20-$30 floor in late August and Cursor's marketing page reflects that, while xAI's own developer-facing FAQ simply has not been updated to match. Anyone deciding whether to subscribe should check both pages directly rather than trust either one in isolation, and should expect the number to keep moving: three changes in fifteen days is not a settled pricing structure, it is a beta finding its floor.

One structural detail holds regardless of which tier unlocks access: Grok Bot usage is metered separately from the rest of a user's SuperGrok or Cursor plan. xAI's own FAQ states it directly: "Grok Bot subscriptions include weekly usage; eligible accounts can add on-demand usage billed from model and token cost," and for users who hold both a Cursor and a SuperGrok subscription simultaneously, "Grok Bot uses whichever has more usage" - xAI Grok Bot FAQ. That is a meaningfully different arrangement from the shared weekly pool described in Section 2 for Chat, Imagine, and Voice. Grok Bot does not draw from that same pool; it has its own, and its own is the one with no published size, which is the subject of the next section.

4. The Real Limits: Weekly Usage Pools, No Spend Cap, and the Bills That Surprised Early Users

The single most important thing to understand about Grok Bot's cost, separate from which subscription unlocks it, is that xAI does not publish how much usage any tier actually includes. A researcher who read through xAI's pricing pages, plan descriptions, and FAQ documentation as of late August 2026 summarized the finding in one sentence: "xAI publishes zero per-tier numbers" - Christopher Alarcon. That is not an oversight in a still-forming product category; every other agent platform in this guide's assessment table (Section 9's Devin, Manus, and Cursor Cloud Agent among them) publishes a specific credit, task-count, or ACU-hour figure a buyer can plan a budget around. Grok Bot, alone among the six, does not.

What xAI does confirm is the mechanism, even without the numbers. Grok Bot subscriptions include a weekly usage allowance, and overage beyond that allowance bills at standard model and token rates, with no dedicated spend cap for the product - eesel AI. There is also no model picker inside Grok Bot: a user cannot route a routine task to a cheaper model the way they might inside a coding tool, so every action, regardless of complexity, draws against the same rate - eesel AI. Combined, those two facts describe a product where cost scales with usage in a way the subscriber cannot see coming and cannot cap in advance, which is a materially different risk than a $20/month chat subscription with a fixed monthly price.

That risk showed up immediately in practice. eesel AI's teardown of the launch pricing quotes two builders discussing real bills: Greg Isenberg and Billy Howell described the tool costing "$200, $300 a month," with Howell specifically noting he paid $200 a month for a second, separate account to keep pace with his workload - eesel AI. A more extreme case came from an early tester who reported consuming more tokens in a single month of always-on Grok Bot usage than in the previous five years of ordinary interactive chat use combined, attributing the spike to routines that poll every 15 minutes (96 times a day, every day) and long-context chief-of-staff-style threads that accumulate history with every message - eesel AI.

• A routine polling every 15 minutes runs 96 times a day, and each run consumes pool budget whether or not it finds anything worth acting on
• Long-running chat threads with a Bot carry accumulating context, so a week-old conversation costs more per message than a fresh one
• Video and image generation inside Grok Imagine draws disproportionately on the same shared pool used for text chat, as covered in Section 2
• No model picker means there is no cheaper fallback for routine, low-stakes actions the way there is in most coding agents

Those four mechanics interact in a way that is easy to underestimate before actually running an always-on agent. A single Bot left checking an inbox every 15 minutes is not, by itself, an expensive workload. The problem compounds when a user runs several Bots in parallel (which the product explicitly supports and encourages, since coordinating multiple Bots is a headline feature) and each one accumulates its own long-running context over days or weeks. What reads as one subscription fee on the pricing page becomes, in practice, a variable cost that scales with how many always-on routines a subscriber sets up and how long each one has been running, with no dashboard showing a hard ceiling until the weekly meter is already close to empty.

The billing reality does have one practical mitigation worth noting for anyone already paying for multiple qualifying plans. Since Grok Bot draws from whichever of a user's plans (Cursor or SuperGrok) has more remaining usage, a subscriber who holds both is effectively pooling two allowances rather than being capped by the smaller one - xAI Grok Bot FAQ. That is a genuine cost lever, though it also means the true entry cost for heavy Grok Bot use may functionally require paying for two separate subscriptions rather than one, undercutting any headline "$20/month" framing for a user who actually wants to avoid running dry mid-week.

It is worth running the inbox-triage example from earlier through actual numbers to see how quickly an apparently modest routine adds up, using the disclosed API rates from Section 6 as the closest available proxy for what the underlying compute actually costs, since xAI does not publish a separate Grok Bot rate card. A routine that polls an inbox every 15 minutes runs 96 times a day, 672 times a week. Even a conservative estimate of 2,000 tokens of context per check (the message thread, the account's routing rules, the model's response) works out to roughly 1.3 million tokens a week for that single routine alone, which at Grok 4.6's disclosed API rate of $2 to $6 per million tokens lands somewhere between $3 and $8 a week just for polling overhead, before the routine has drafted a single reply or taken a single action. Multiply that by three or four Bots running similar routines in parallel, which the product explicitly encourages, and a subscriber can be several tens of dollars into token spend in an ordinary week without a single unusually large task having run at all. That arithmetic is the mechanical explanation behind the tester's five-years-of-usage-in-one-month report cited above: it is not one expensive task, it is many small ones compounding continuously, 24 hours a day, in a way an interactive chat habit never does.

• Estimate polling frequency first. A routine checking every 15 minutes costs roughly 8 times more in baseline overhead than one checking hourly, for identical per-check token cost
• Count concurrent Bots, not just tasks. Three Bots each polling modestly compound faster than one Bot doing more per check
• Watch context growth in long threads. A week-old conversation with a Bot costs more per exchange than a fresh one, since full history is typically resent
• Treat Grok Imagine as a separate budget line, not an occasional extra, given its disproportionate draw on the shared pool

None of those four levers are visible in a subscription's advertised price, and none of them show up until the weekly usage meter is already most of the way to empty. That is the practical, bottom-line reason this guide treats Grok Bot's usage transparency as the single lowest-scoring criterion in Section 9's assessment table: the cost is genuinely knowable in principle, using the same API math worked through above, but xAI does not do that arithmetic for the subscriber, and no dashboard inside the product currently breaks a weekly bill down by routine, by Bot, or by task type the way a cloud billing console would.

5. Security and Failure Modes: The Shared-Computer Problem

Cost is not the only risk Grok Bot introduces, and the security model is directly relevant to a pricing guide because it shapes how a rational buyer should size their usage: running fewer, more carefully scoped Bots is both a cost control and a security control here. xAI's own documentation is unusually direct about the limitation, more direct than most vendors are about their own agent products. The Grok Bot FAQ states plainly: "Do not use separate Bots as a security boundary" - xAI Grok Bot Security & Privacy.

The underlying architecture explains why that warning exists. Every Bot a user creates runs on one shared, persistent cloud computer assigned to that user's account, and every Bot on that computer shares its files, browser sessions, and command-line credentials - xAI Grok Bot Security & Privacy. The isolation xAI does guarantee is between users, not between a single user's own Bots: one person's Grok Bot computer cannot reach another person's, but within one account, a login or file created by one Bot is available to every other Bot that account runs. The documentation frames each Bot's screen as "a separate work surface, not a separate security boundary," and its own recommended mitigation for anyone who genuinely needs isolated credentials is blunt: create an entirely separate Cursor or SuperGrok user for that workload, not a separate Bot under the same one - xAI Grok Bot Security & Privacy.

Because the computer holds real, standing credentials rather than scoped, revocable tokens, the practical consequence of a single compromised Bot is that the entire account's access is exposed, not just the one workflow that Bot was assigned. That architecture puts Grok Bot squarely inside a risk category the industry has already learned to worry about from a separate, earlier incident. In May 2026, an attacker drained roughly $150,000 to $174,000 in DRB tokens from a crypto wallet that Grok had auto-provisioned through its interaction with Bankr, a third-party trading bot on X, using a three-step attack: first sending an NFT that escalated Grok's wallet permissions, then asking Grok to translate a Morse code message that decoded to a transfer instruction, which Grok's safety filter read as a harmless translation request rather than a command - Giskard. That incident predates Grok Bot by three months and involved a different product (Grok interacting with a third-party wallet bot, not the agent platform itself), but it demonstrates the exact failure class Grok Bot's shared-credential architecture inherits: an encoded or otherwise disguised instruction reaching an agent that holds real permissions, bypassing keyword-based safety filters because the malicious content never appears in plain text.

Industry-wide data suggests this is not a hypothetical risk specific to one vendor. OWASP's GenAI Security Project ranked agent goal hijacking as the single highest-priority risk in its 2026 State of Agentic AI Security report, and controlled experiments in that research found credential-theft outcomes in roughly 70% of prompt-injection trials run against agentic systems - Giskard. Grok Bot's architecture, real standing credentials on a computer shared across every Bot a user runs, is not unusually reckless by the standards of this product category. It is representative of a category-wide problem that the whole computer-use agent space, not just xAI, is still working out, and O-mega's own coverage of prompt-injection defense in agentic systems goes deeper into the mitigation patterns available across vendors - AI Agent Security: Prompt Injection Defense.

xAI's mitigations, as documented, are procedural rather than architectural: an approval workflow requires explicit confirmation before actions like publishing, purchasing, deleting, or changing permissions, and the interface explicitly refuses to accept passwords, one-time codes, or CAPTCHAs typed directly into chat, routing those instead through a manual entry step - xAI Grok Bot Security & Privacy. On desktop, a pending action surfaces an "Allow once," "Deny," or "Always allow" choice, with the mobile app offering the same decision as "Approve once" and "Deny"; an administrator can also configure Auto Review rules that either always stop a matching action for manual review or always allow it, with a "Require Approval" rule taking precedence whenever it conflicts with an "Always Allow" rule covering the same action type - xAI Grok Bot Security & Privacy. One limitation of that approval layer is stated in the documentation itself and is easy to miss: "an approval controls the proposed action, it does not reverse work already completed," meaning the gate only stops a Bot before it acts, not after, which is no protection at all against an action that completes faster than a human reviewer can respond to a prompt.

Those controls reduce risk for the specific action they gate, but they do not change the underlying fact that the computer itself is a single shared trust boundary per user. For a buyer weighing whether the cost of Grok Bot is worth it, that single fact should shape which workflows get assigned to it: routine, low-privilege tasks are a reasonable fit, while anything touching financial credentials, customer data, or production systems is exactly the kind of workload xAI's own documentation says a separate Bot cannot safely isolate.

6. The Developer API: Grok 4.6 Token Pricing for Builders

Separate from every consumer and agent tier covered so far, xAI sells direct API access to its Grok models, billed per million tokens the way every other frontier lab prices its developer product. As of this writing, the current flagship available through the API is Grok 4.6, described in xAI's own model documentation as "the most intelligent and fastest model" in the lineup for general chat, code, and agentic use - xAI Models. It sits alongside several other model variants still live in the same API, including the prior-generation Grok 4.5, the cheaper Grok 4.3, a specialized grok-build-0.1 model (the same "Build" product line that underlies Grok Bot's own compute pool), and a dedicated grok-4.20-multi-agent-0309 variant built for parallel multi-agent workloads.

Source: xAI's live models documentation, fetched September 2026. Pricing roughly doubles for requests exceeding the 200k-token threshold on each model.

The pricing structure carries one detail worth calling out explicitly for anyone budgeting a production workload: cost scales with context length in a step function, not gradually. A request under 200,000 tokens on Grok 4.6 costs $2/$6 per million tokens; the identical model handling a request that crosses that 200k threshold jumps to $4/$12, a full doubling rather than a linear increase - xAI Models. That is a meaningfully different pricing shape from competitors who price context length on a smoother curve, and it means a workload that occasionally spikes past 200k tokens (a long document analysis, an extended agent trace) pays a step penalty on the entire request, not just the portion beyond the threshold.

A concrete comparison shows how much that step function actually costs a real workload. Processing a 50-page legal document, roughly 40,000 tokens once converted to text, through Grok 4.6 costs about $0.08 for the input side at the sub-200k rate. Extend the same task to a 300-page filing, closer to 240,000 tokens, and the request crosses the 200k threshold entirely, so the full 240,000 tokens bill at the higher $4-per-million rate rather than only the portion past the line, bringing input cost to roughly $0.96, a twelvefold increase for a document only six times longer. Builders who batch-process long documents have a direct incentive to chunk inputs to stay just under 200,000 tokens per call wherever the task allows it, since the pricing structure rewards that discipline far more aggressively than a smoothly scaling per-token model would.

For teams evaluating whether to build directly on the API versus paying for a SuperGrok or Grok Bot subscription, the calculation genuinely depends on usage pattern. A predictable, moderate-volume application (a support chatbot handling a known number of daily conversations, for instance) is usually cheaper and more controllable on the API, since the cost per request is fixed and disclosed, unlike the opaque weekly pools covered in Sections 2 and 4. An unpredictable, spiky workload, or one where a non-technical team member needs to interact with Grok directly rather than through custom-built infrastructure, is where the subscription tiers make more sense despite their usage-visibility problems. Teams optimizing multi-model spend across several providers, not just xAI, may find more value in routing strategies than in committing to any single vendor's ladder, a pattern covered in more depth in O-mega's guide to cutting AI agent costs with model routing, and the underlying token-cost mechanics behind that decision are broken down further in The True Cost of LLM Inference in 2026.

7. Grok for Business and Enterprise: Per-Seat Pricing for Teams

Separate again from the consumer ladder and the developer API, xAI runs a dedicated team product with its own pricing and, notably, its own data-handling policy. Grok Business costs $30 per seat per month and includes self-serve admin controls, SOC 2 Type II compliance, role-based access control, a 128K context window, and consolidated billing across a team's seats - Medianama.

The single most consequential difference between Grok Business and every consumer tier in Section 2 is a training-data policy, not a feature. By default, xAI does not train its models on data submitted through Grok Business accounts, while free, X Premium, and individual SuperGrok plans may be used for that training unless a user separately opts out - Medianama. For any organization handling client data, proprietary code, or anything else that should not risk surfacing in a future model's training set, that distinction alone often justifies the jump from a $30/month individual SuperGrok plan to a $30-per-seat Business plan, even though the headline number looks identical.

Above Business sits Grok Enterprise, custom-quoted rather than published, adding custom single sign-on, SCIM directory sync, more granular role-based access control, advanced audit and security controls, dedicated onboarding support, and an "Enterprise Vault" data-handling tier - Medianama. That tier structure (a transparent, self-serve mid-market plan below an opaque, sales-assisted enterprise plan) mirrors how every other major AI lab structures its business offering, including Anthropic's Team plan and OpenAI's own Business tier. What is specific to Grok's version is the emphasis on the training-data guarantee as the headline reason to upgrade, reflecting xAI's awareness that data provenance, not raw model capability, is often the actual blocker preventing a company from adopting any consumer-grade AI tool at all.

Set side by side with its two closest peers, Grok Business's $30-per-seat price sits almost exactly between Anthropic's Claude Team Standard seat at $20 to $25 per month and its Premium seat at $100 to $125, and roughly in line with OpenAI's Business Standard seat at $25 - Claude. None of the three vendors compete primarily on that headline number, since a $5 to $10 per-seat difference rarely decides a procurement conversation on its own. What actually differs between them is the specific bundle behind the price: Grok Business's pitch centers on the default no-training guarantee and real-time X data access, Claude's centers on Claude Code and Claude Cowork shipping in the same seat that also covers ordinary chat, and OpenAI's centers on breadth of integrations across the wider Microsoft and enterprise software ecosystem. A team choosing between them is realistically choosing which existing workflow and which data-handling story fits best, not which seat is a few dollars cheaper.

8. Alternatives: AI Chatbot Subscriptions

For a reader whose actual need is a general-purpose chat assistant rather than an always-on agent, Grok's $30/month SuperGrok tier competes directly against three other frontier labs' equivalent offerings, and the honest comparison depends heavily on what a subscriber actually does with the tool day to day rather than any single benchmark number. Anthropic's current consumer lineup runs Claude free at $0, Claude Pro at $17/month billed annually or $20/month billed monthly, and Claude Max starting at $100/month for five times the Pro usage allowance or $200/month for twenty times - Claude. Notably, Claude Pro already bundles both Claude Code and Claude Cowork (Anthropic's own always-on desktop agent, covered again in Section 9) into the same $17-$20 subscription, which is a materially different bundling decision than xAI's, where Grok Bot access sits behind a separate, higher tier entirely.

OpenAI's ladder runs Free (unlimited basic text chat), Go at $8/month, Plus at $20/month, and Pro at either $100 or $200/month depending on a usage-multiplier toggle on the pricing page itself - AI Pricing Guru. The company's current flagship reasoning model, GPT-6 Astra, began rolling out to approved organizations on September 3, 2026 and to all ChatGPT Plus, Pro, Business, and Enterprise users within days, with the rollout notes specifying that "Astra usage is included within the existing subscription allowances," meaning the upgrade did not come with a price increase attached - 9to5Mac. Google's Gemini access follows a similar shape through Google AI Pro at $19.99/month and two tiers of Google AI Ultra at $99.99 or $199.99/month for 5x or 20x the Pro usage ceiling respectively, bundled with substantial cloud storage and Google Workspace AI features that neither xAI nor OpenAI offer as part of the same subscription - Google AI Premium. As of this writing, Google's best publicly available reasoning model is Gemini 3.1 Pro, while its flagship Gemini 3.5 Pro remains delayed past its originally announced mid-2026 window - Fortune.

Perplexity occupies a different niche entirely, built around search and citation-grounded answers rather than open-ended conversation, and its pricing reflects that specialization: Pro runs $20/month for the standard research-and-search product, while Max at $200/month adds Perplexity Computer, a system that orchestrates 19 different specialized sub-agent models rather than a single general chatbot - Finout. Perplexity licenses its underlying models from OpenAI, Anthropic, and Google rather than training its own frontier model, which is a structurally different bet than every other row in this table: it is not competing on raw model capability, it is competing on retrieval quality and citation accuracy layered on top of whichever model performs best for a given query. Enterprise buyers evaluating Perplexity face the same seat-based structure as the rest of this field, with Enterprise Max priced at $325 per seat monthly or $3,250 annually, positioned well above every other enterprise seat covered in Section 7 and reflecting the cost of licensing three separate labs' frontier models simultaneously rather than training one in-house - Perplexity AI Magazine.

The free tier across all five products deserves a direct comparison too, since it is where most new users actually start before ever seeing a pricing page. Meta AI remains free with no paid consumer tier as of this writing, built into Instagram, WhatsApp, and Messenger rather than sold as a standalone subscription, which makes it structurally incomparable to the other five on a per-dollar basis but relevant for anyone weighing whether a paid Grok subscription is worth it against a free alternative already embedded in apps they use daily. Grok's own free tier, by contrast, is the most usage-limited of the group by design, running the lighter Grok 4 Mini model with no image generation at all, a deliberate wedge intended to push casual users toward at least the $8 X Premium tier rather than let them settle indefinitely on the free plan the way ChatGPT's unlimited-basic-chat free tier allows.

Set against that field, SuperGrok's positioning is clearer than it first appears. At $30/month it costs 50% more than the $20 default tier every other major lab charges, and the honest reason to pay that premium is real-time access to X's public conversation firehose through Grok's Live Search and DeepSearch features, a data source none of the other four products can match natively, rather than any clear edge in general reasoning benchmarks. A reader who wants a chatbot for writing, coding help, or research is generally better served checking O-mega's dedicated benchmark comparisons for Grok 4.5 and Grok 4.6 against the current field, or the standalone Claude Sonnet 5 benchmark breakdown, rather than treating "which chatbot is cheapest" as the deciding question in isolation from what each model is actually good at. Platforms like O-mega take a different approach entirely from any single-model subscription: rather than picking one lab's chatbot, O-mega builds and runs an entire autonomous company across whichever underlying models fit each task, Grok included, so the model-selection question becomes something the platform handles per task rather than something a subscriber locks in for a flat monthly fee.

9. Alternatives: Always-On Agent Platforms

The more consequential comparison, given everything Sections 3 through 5 documented about Grok Bot's cost opacity and security model, is against the other products actually competing in the same always-on, computer-operating agent category. The scoring table at the top of this guide ranked six of them; this section fills in the reasoning behind each score with the specifics that did not fit in a table cell.

ChatGPT Agent, OpenAI's evolution of its earlier Operator product, scored highest in this guide's assessment specifically because of how disclosed its limits are. A ChatGPT Plus subscriber gets 40 agent tasks per month included in the $20 subscription, and additional usage is priced transparently at roughly $50 for 100 tasks, about $0.50 per task - AI Pricing Guru, a full breakdown of which is available in O-mega's dedicated ChatGPT Operator pricing guide. That is the exact opposite of Grok Bot's undisclosed weekly pool: a ChatGPT Agent user can calculate, in advance, exactly what a month of heavy usage will cost. The tradeoff is scope. ChatGPT Agent operates primarily through a sandboxed virtual browser rather than a full desktop computer with terminal and filesystem access, which narrows what kinds of workflows it can automate relative to Grok Bot's broader reach, even as GPT-6 Astra's September 2026 rollout specifically targeted faster, more capable computer use as one of its headline improvements - 9to5Mac.

Claude Cowork, Anthropic's own always-on desktop agent, takes a bundling approach unlike either OpenAI's or xAI's: it ships inside the same $17-$20/month Claude Pro subscription that also includes Claude Code, rather than sitting behind a separate, higher-priced tier the way Grok Bot does relative to base SuperGrok - Claude. That pricing decision alone makes Claude Cowork the cheapest entry point into always-on agent capability among the frontier labs, though its usage is still governed by the same 5x/20x multiplier structure that Claude Max applies to ordinary chat, rather than a separately disclosed agent-specific allowance. O-mega's own coverage of Claude Cowork's pricing and agent ecosystem and its growing list of Claude Cowork alternatives is worth reading in full for anyone specifically comparing it against Grok Bot feature by feature.

Devin, from Cognition, remains the most narrowly scoped but most mature entrant here, built specifically for software engineering rather than general computer use, a distinction O-mega weighed directly in Devin Desktop vs Claude Code: The 2026 Guide. Its pricing is also the most granular: a Core plan at $20/month plus $2.25 per Agent Compute Unit, where one ACU represents roughly 15 minutes of agent work, putting an hour of sustained autonomous coding at around $9 - VP0 Journal. A Team tier adds $500/month for 250 included ACUs, and Max at $200/seat raises the ceiling further for heavier individual users - Automation Atlas. Because Devin's scope is deliberately narrow (pull requests and defined engineering tasks, not arbitrary app control), its reliability track record is the strongest of the six platforms in this comparison, a direct result of not trying to generalize across every possible workflow the way Grok Bot does.

Manus and Cursor's own Cloud Agent (formerly Background Agents) round out the field from opposite ends of the scope spectrum. Manus is broad and research-oriented, running on a credit system that starts with 300 free daily credits and scales through Pro tiers at $20, $40, and $200 per month depending on task complexity and concurrency needs - Fello AI. Cursor's Cloud Agent, by contrast, is deliberately narrow: it is not the same product as Grok Bot despite living inside the same company's pricing page, and xAI's own documentation is explicit about when to use which, a distinction that matters for anyone weighing Cursor generally against the other entrants in O-mega's Claude Code vs Codex vs Cursor comparison. A Cloud Agent is the right tool "when the output should be a branch and a PR inside an isolated VM," while Grok Bot is the right tool "when the job lives in apps, files, a browser, and a computer that stays on after the laptop closes" - Continuum Code. That distinction, scoped isolated task versus persistent shared computer, is worth remembering for anyone comparing the two products just because they share a vendor and a pricing page.

• ChatGPT Agent: best transparency, browser-scoped, $20/month for 40 tasks
• Claude Cowork: cheapest bundled entry at $17-$20/month, governed by Claude's usage multiplier
• Devin: narrowest scope, strongest track record, granular per-ACU billing around $9/hour
• Manus: broadest research and browsing scope, credit-metered with published costs
• Cursor Cloud Agent: code-only, isolated VM per task, included in the $20/month Pro plan

Reading across those five alongside Grok Bot itself, the pattern that emerges is not that any single platform is simply better across the board, it is that each one made a different tradeoff between scope, price, and transparency, and Grok Bot is the only one of the six that maximized scope (full desktop computer use, any app, no API required) while minimizing transparency (no published allowance, no spend cap, no model picker) at the same time. That combination is exactly why it scored lowest on the two heaviest-weighted criteria in this guide's assessment table despite genuinely leading the field on raw capability.

It is worth being explicit about what "leading on raw capability" actually means here, since it is easy to read as a vague compliment rather than a specific, checkable claim. Among the six platforms compared, Grok Bot is the only one that requires no API integration, no MCP connector, and no purpose-built plugin to operate a third-party tool: it drives the same browser, the same login screen, and the same buttons a human employee would use, which is why xAI can credibly promise it works with software that has never published an API at all - x.ai official announcement. Devin and Cursor's Cloud Agent, by contrast, operate almost entirely inside version-controlled code repositories where an API-shaped interface (git, a test runner, a CI pipeline) already exists by construction. ChatGPT Agent and Manus sit in between, using a controlled virtual browser rather than either a full native desktop or a code-only sandbox. That specific difference, native desktop control versus browser sandbox versus code repository, is the real axis this category is organized around, and it explains why no single platform tops every criterion in the assessment table: broader control surface area consistently trades against tighter, more auditable security boundaries, and no vendor in this comparison has yet shipped both at once.

10. How to Choose: A Decision Framework by Use Case

The right Grok product, or the right alternative to it, depends almost entirely on which of three distinct jobs a reader actually needs done, since the pricing and risk profiles covered in this guide diverge sharply between them.

For someone who wants a general chat assistant, real-time X data access, and nothing more, the decision is straightforward: X Premium at $8/month is the floor for any Grok access at all, and SuperGrok at $30/month is the right upgrade the moment DeepSearch, Voice mode, or Grok Imagine become part of the regular workflow. That user should not touch Grok Bot at all; the agent product solves a different problem and introduces the security and billing exposure covered in Sections 4 and 5 for no benefit to someone who only wants to chat. Anyone in this category who is specifically choosing between Grok and a competing chatbot on capability rather than price alone should treat this guide's pricing numbers as one input among several, and weigh them against actual benchmark performance on the tasks that matter most day to day, since the cheapest tier is a false economy if it cannot reliably do the work.

A third group sits between those two and is easy to overlook: someone already paying for SuperGrok or a Cursor plan for unrelated reasons, who is deciding whether to simply try Grok Bot rather than actively shop for an agent platform from scratch. For that reader, the marginal cost of experimenting is genuinely low, since Cursor's own current pricing page places basic access inside the $20 Pro plan many of them already hold. The right frame for that trial is not "is this worth $20 a month" (a question already answered by the existing subscription) but "which of my own recurring, low-privilege tasks would I actually trust to an unattended agent with no published spend cap," and starting with exactly one narrow, reversible routine such as a daily summary rather than anything touching money, customer data, or production systems is the sizing discipline every source in Sections 4 and 5 points toward.

For someone specifically evaluating an always-on agent to operate real software unattended, the calculation is less about price and more about risk tolerance and workflow scope. A team whose need is narrowly scoped to software engineering tasks is almost certainly better served by Devin's predictable per-ACU pricing or Cursor's scoped Cloud Agent than by Grok Bot's broader but far less transparent computer-use model. A team that genuinely needs an agent operating across arbitrary business software (a CRM, an inbox, an internal dashboard with no API) is the specific use case Grok Bot was built for, and the tradeoff to accept going in is that neither the weekly cost ceiling nor the credential-isolation boundary between that team's own workflows is fully under their control yet, both being genuine, documented, current limitations rather than hypothetical ones.

For a builder or founder deciding how to run an entire operation rather than subscribing to any single vendor's chat or agent product, the more durable answer is usually not "which one subscription" but "which orchestration layer lets me use the right model for each task without re-platforming every time a new one ships." That is the problem O-mega is built to solve, running browser automation, computer automation, and internal agent sessions across whichever underlying models fit best, Grok's included, rather than locking an entire company's workflow into one lab's pricing ladder and one lab's usage-transparency choices.

Grok's 2026 pricing story, in the end, is really two stories wearing the same brand name. The consumer chat ladder is mature, well-documented, and easy to reason about: seven tiers, a shared weekly pool since June, real prices a subscriber can check against xAI's own pages with confidence. Grok Bot is the opposite: a genuinely capable, genuinely new product still finding its price fifteen days at a time, with a usage meter its own vendor will not fully disclose and a security model its own documentation warns about in plain language. Both are real. Neither should be assumed to still be true by the time this guide is six months old, which is exactly why every figure in it links back to the primary source it came from rather than asking for trust.

This guide reflects Grok, SuperGrok, X Premium, and Grok Bot pricing as of September 2026. xAI changed Grok Bot's access tiers three times in the first fifteen days after launch and has not published exact usage-pool sizes for any tier; verify current prices and eligible plans directly on xAI's and Cursor's own pricing pages before subscribing, since even those two primary sources did not fully agree with each other at the time of writing.