Grok

Grok is xAI’s general-purpose AI assistant, available on grok.com, iOS and Android, and separately through X. The standalone product now combines chat, real-time web and X search, voice, file and image analysis, connectors, image/video generation through Imagine, and increasingly agentic work. xAI’s current product documentation describes Grok as a synchronized web/mobile assistant rather than merely a chatbot embedded in X.¹

The model layer and the consumer product need to be separated. Grok 4.6 launched on August 12, 2026 as xAI’s current frontier model for coding, long-running agents and knowledge work. xAI’s individual pricing page explicitly includes Grok 4.6 in the $30/month SuperGrok tier. The free tier is described as free Grok access with generous limits, but the current pricing page does not promise that every free chat is routed to Grok 4.6.²³

That distinction controls this review. AiToolMap uses Grok 4.6 benchmark evidence as a direct match for the SuperGrok model entitlement and as evidence about xAI’s current technical ceiling. It does not silently transfer those benchmark scores to every free or X-hosted Grok interaction.

Artificial Analysis is the strongest current independent model benchmark in the fixed panel because the exact Grok 4.6 page was opened and read. Grok 4.6 High scores 61 on the Intelligence Index and ranks #6/186 in the captured current class. The same page reports $2 per million input tokens, $6 per million output tokens, a 500k context window, about 65.8 output tokens per second and a high time-to-first-token of roughly 41.9 seconds.⁷

The profile is therefore not “cheap and fast.” It is better described as high capability at competitive token economics, with latency as a real trade-off. The high reasoning setting is slower than the median comparison model and can be verbose. A user choosing Grok because of the 61 index score should expect that the strongest setting may feel materially less instantaneous than lightweight chat modes.

xAI’s first-party launch material adds task-level detail. Grok 4.6 improves over 4.5 on coding, knowledge-work and agentic benchmarks and is specifically trained for long-running multi-step work. xAI also reports broader pre-deployment safeguard testing for this generation.² Because these are vendor-reported scores, AiToolMap uses them as product/model documentation rather than independent validation.

The consumer-routing qualification matters. SuperGrok explicitly lists “Grok 4.6 model” among its benefits.³ The free plan does not make the same explicit promise. The App Store product copy also still describes Grok 4 for difficult math, science and coding tasks in the captured listing.⁸ This may lag the live product, but it is another reason not to assume one universal model route across every user and surface.

The current standalone Grok product covers the major assistant surfaces: normal conversation, web/X retrieval, voice, uploads, connectors and media creation.¹ It is particularly differentiated where up-to-the-minute public information matters because web search and X search are core product features rather than add-ons.

That advantage is not identical to factual reliability. Real-time access can improve freshness while still importing noisy public information. xAI’s own FAQ warns that Grok can hallucinate or return misleading information and recommends verification.⁵ The correct framing is therefore “strong current-information access,” not “truth by virtue of X access.”

Voice is built into the free product according to current pricing, and both mobile stores present Grok as a broad everyday assistant.³⁸⁹ Imagine adds integrated image and video creation. SuperGrok Plus further adds 1080p video and Grok Bot access, moving the product toward persistent AI work rather than only conversation.³¹⁶

For users who want one subscription spanning reasoning, live research, voice and creative generation, Grok’s bundle is unusually broad. Users who primarily need Microsoft 365 or Google Workspace-native workflow depth may still find Copilot or Gemini structurally better integrated into their existing environment.

xAI launched Grok Bot as an AI teammate running on a persistent cloud computer and expanded access in August to SuperGrok Plus and selected Cursor plans.¹⁶ This is strategically important because xAI is pushing beyond “answer this prompt” into longer-running delegated work.

The evidence is still too young to call Grok Bot mature. AiToolMap found current first-party positioning but not a robust independent panel of long-duration Bot reliability tests. The review therefore treats agentic breadth as a capability strength and maturity as unproven rather than awarding assumed reliability.

This caution is reinforced by product-security evidence elsewhere in the Grok family. The Register’s August report describes a cryptographic prompt-injection attack against Grok.com that could cause injected instructions to be followed when a user asked Grok to summarize a malicious page.¹³ That is directly relevant to any future in which Grok receives more permissions to browse, read and act.

xAI’s current individual pricing page is unusually clear at the headline level.³

**Free — $0/month.** Includes Grok within usage limits, real-time web/X search, voice, connectors and SOC 2 Type I & II statements on the pricing page.

**SuperGrok — $30/month.** Explicitly adds Grok 4.6, higher rate limits across features, Expert, connectors and image/video generation.

**SuperGrok Plus — $100/month.** Adds substantially higher usage, 1080p video creation, priority access, faster replies and Grok Bot.

X subscriptions are a separate route. X Premium/Premium+ can provide higher Grok limits and some account combinations include SuperGrok, but Grok on X is governed by X terms/privacy and should not be silently collapsed with the standalone product.²⁰⁶

For developers, Grok 4.6 costs $2 per million input tokens, $0.50 per million cached input tokens and $6 per million output tokens in the standard context tier, with a 500k context window.²¹ Artificial Analysis’s independent $0.84 estimated Intelligence Index task cost supports the argument that the model’s capability/cost ratio is competitive even though high reasoning is not especially fast.⁷

At $30, SuperGrok is more expensive than the familiar $20/month mainstream assistant tier. Its value case is strongest when a buyer genuinely uses the combination of Grok 4.6, live X/web search, voice and media generation. If the buyer mainly wants long-form writing or conventional chat, the premium is harder to justify.

SuperGrok Plus at $100 is not simply “better answers.” It is a capacity/agent/media tier. The economic case depends on heavy use of Bot, higher media quality, priority access and higher overall limits.³

The enormous app-store populations are difficult to dismiss. A 4.9/5 U.S. App Store rating across roughly 1.3 million ratings and a 4.8/5 Google Play rating across roughly 3.67 million reviews imply broad consumer satisfaction at scale.⁸⁹ They do not tell us which model route each reviewer used, but they are valuable product-level evidence.

Current editorial evidence supplies a sharper failure mode. On August 20, TechCrunch reported that some Grok Lite users received long, nonsensical strings of English after normal requests, including a PDF-related task. TechCrunch could not reproduce the failure and judged it likely limited to a subset of users; starting a fresh session often restored normal behavior.¹²

That scope matters. It would be wrong to describe Grok generally as “broken.” It would also be wrong to ignore a documented current reliability regression because aggregate app-store ratings are high. AiToolMap therefore treats this as a product reliability caveat specifically tied to Grok Lite/Grok.com in the observed period.

Trustpilot points in the same qualitative direction but is not numerically integrated. The current Grok page is heavily negative and includes complaints about glitches, limits and paid-service friction, but the population is small and self-selected. Under methodology v1.8 Trustpilot is used only to identify complaint themes, never as a direct rating vote.

G2’s current Grok page reports 4.5/5 from 141 reviews. Its synthesis highlights fast research, real-time X integration, ideation and responsiveness, with some users saying information depth can require follow-up prompting.¹⁰

Gartner Peer Insights is more cautious: 3.7/5 from three ratings. One August enterprise reviewer says the expert model generated accurate network configurations that worked but was slow; another review criticizes logic, complex formulas, coding and free-tier restrictions.¹¹ The sample is too small to compete statistically with millions of mobile ratings, but it is useful as a reminder that professional/workflow satisfaction can differ from consumer app enthusiasm.

Capterra and TrustRadius both had current Grok pages discoverable in this audit, but the underlying pages could not be substantively opened in the retrieval workflow. Their search-result ratings are therefore excluded from the v1.8 numeric inputs. PeerSpot’s underlying substantive review is from March 2025 and is excluded as stale. This absence discipline is intentional.

Current safety evidence is not limited to abstract policy debates. Reuters reports that Canada’s privacy commissioner concluded Grok’s image-generation tool launched without appropriate safeguards against non-consensual sexualized deepfakes; xAI later committed to stronger monitoring and controls.¹⁴

WIRED independently inspected hundreds of public Grok Imagine links and found dozens leading to sexualized AI images or videos, including non-consensual depictions of real women. The article notes that xAI added safeguards and removed material after contact, while experts and regulators questioned their sufficiency.¹⁵

The web-agent risk is different. The Register reports that encoded malicious instructions could still manipulate Grok.com in August after xAI had been alerted, illustrating the familiar prompt-injection problem for assistants that browse external pages.¹³

The same publication and Simon Willison also documented the separate Grok Build controversy: the coding CLI had uploaded repository/directory contents to xAI-controlled cloud storage before xAI disabled the behavior, promised deletion and open-sourced the CLI.¹³¹⁷ This should not be attributed to ordinary chat, but it matters when assessing trust across xAI’s expanding agent ecosystem.

Government evidence is likewise mixed. The Wall Street Journal reports that multiple U.S. agencies raised concerns about Grok safety and reliability in sensitive deployments, with a GSA assessment saying Grok-4 did not meet the safety/alignment expectations required for general federal use without strict oversight.¹⁸ Reuters separately found limited federal and enterprise adoption relative to leading competitors in its May review of government inventories and Netskope corporate usage data.¹⁹

These reports do not prove that a normal consumer should avoid Grok. They do show why a high Artificial Analysis score should not be treated as a complete trust score.

xAI’s current standalone-Grok privacy policy is effective August 24, 2026. It applies to Grok.com and the standalone iOS/Android apps; Grok accessed through X is governed by X’s separate privacy framework.⁴

The policy says xAI can collect account information, user content, feedback, technical/usage data, location-related data, public X posts and internet-search data depending on how the service is used. It also says personal information can be used to develop and improve the service, including model training.⁴

The consumer FAQ gives users a material control: on standalone Grok web/mobile, users can turn off “Improve the model.” New conversations after opting out are not used for training, except that voluntarily submitted feedback may still be used. Private Chat, where available, is not used for model training.⁵

Private Chat also changes retention. The current privacy policy says Private Chat conversations do not appear in normal history and are deleted from xAI systems within 30 days unless longer retention is necessary for legal, compliance or safety reasons. Deleted normal conversations/accounts are likewise scheduled for deletion within 30 days subject to those exceptions.⁴

For European users, xAI points to a Europe-specific privacy addendum. Users should also distinguish standalone Grok from Grok on X: a privacy choice in one surface should not be assumed to control the other without checking the corresponding settings.⁴⁵

Grok is a strong choice for users who want a top-tier model together with very current public information, especially X-native context. It is also compelling for people who use voice and media creation heavily and prefer one assistant that spans research, conversation and creative work.

SuperGrok makes more sense than the free tier for buyers specifically selecting Grok because of Grok 4.6. The $30 price buys an explicit model entitlement rather than merely higher generic usage.³

SuperGrok Plus is for heavy users who can exploit Grok Bot, higher media quality and substantially more capacity. It is difficult to justify solely as a conventional chatbot subscription.

Organizations handling sensitive data or granting agentic permissions should review the current security/privacy terms and test their own workflows rather than relying on consumer app ratings or model benchmarks. Recent prompt-injection and data-handling incidents make that diligence particularly important.¹³⁴