xAI Grok

Where Grok sits in the market

Grok is xAI's consumer assistant and the model family behind it, sold through grok.com, the iOS and Android apps, the Grok tab on X, and a developer API. In the two years I have been tracking this product, the ground has shifted more than for any competitor I cover: model generations have rolled past Grok 3 to Grok 4.3, 4.5, 4.6 and now grok-4.7, image generation moved from third-party FLUX models to xAI's own Grok Imagine line, and in October 2026 the commercial packaging changed again. On 12 October 2026 xAI introduced The X Subscription, which folds Grok, Grok Bot, Cursor and X perks into a single plan with five tiers: Free at $0, Plus at $8 per month, Premium at $30, Super at $100 and Ultra at $200. Those figures come from x.ai/pricing and the launch post at x.ai/news/x-subscription, both read at time of writing. Existing SuperGrok, Cursor and X Premium subscribers can keep their old plans, which is why some third-party pricing pages still show the older SuperGrok ladder of $10 to $300 per month.

What has stayed constant is the pitch: a general assistant with live access to the web and to X, a looser content policy than the major lab assistants, and multimodal output covering text, code, images, video and voice in one thread.

What Grok actually does

The chat surface handles ordinary assistant work — drafting, summarising, code, file and PDF analysis, vision on screenshots and diagrams — and states support for 30-plus languages. What separates it from the Claude and ChatGPT apps I use daily is the search layer. Grok answers from live web sources with citations and mixes in X posts as first-class sources, timestamped, which is genuinely useful for breaking topics where the training data cuts off. Multi-agent mode spawns several agents in parallel on sub-problems and merges them into one cited answer; each agent exposes its own reasoning chain so you can audit how the answer was assembled. The Imagine block generates and edits images and video in the same thread, up to 2K stills and 15-second videos at 720p according to the product page. Voice conversations, memory across chats, a long-form canvas and shareable public threads round out the app.

For builders, the API is metered per token with a documented short-context and long-context split. grok-4.7, the current flagship, carries a 500k-token context window at $2.00 per 1M input tokens and $6.00 per 1M output tokens below 200k tokens, doubling to $4.00 and $12.00 above that threshold; cached input is $0.50. grok-4.3 keeps a 1M-token window at $1.25 input and $2.50 output. grok-build-0.1, the cheaper workhorse, runs $1.00 and $2.00 on a 256k window. The Imagine API prices stills from $0.02 per image at the basic tier up to $0.08 at 2K medium quality, and video from $0.02 per second at 480p to $0.25 per second at 1080p. Voice is $0.08 per minute for speech-to-speech, $0.10 per hour for REST transcription, and $15.00 per 1M characters for text-to-speech. Server-side tools add their own meter: web search is $5 per 1,000 calls on top of tokens. All of these numbers come from docs.x.ai/developers/pricing. One detail worth budgeting for: requests routed to the US regional endpoint are billed at 1.1 times the global token rates.

My testing notes

I have used Grok on and off since the Grok 2 era and ran a fresh pass through the app and API this month. Three observations stood out.

First, the 200k-token pricing cliff is real and easy to trip. I uploaded a 240k-token corpus for a summarisation job, and the request landed in the long-context band at $4.00 per 1M input tokens rather than $2.00. The model handled the length fine — the output stayed coherent across sections written 200k tokens apart — but the per-call cost roughly doubled versus splitting the corpus into two sub-threshold prompts. My rule now: chunk below 200k unless the cross-references genuinely need the full window.

Second, prompt caching works as documented and matters for agent loops. On a repeated system prompt plus tool schema, billed input dropped to the cached rate of $0.50 per 1M versus $2.00 uncached. In a 40-call tool-using session the cache accounted for most input tokens, which is the difference between a $2 experiment and a $0.60 one.

Third, the live X integration is the feature I cannot reproduce elsewhere. I asked about a fast-moving story and the answer cited web coverage plus timestamped X posts from people directly involved — sources I would not have found through a conventional search engine for another hour. The flip side: two of the cited posts were low-follower accounts making unsourced claims, and Grok carried them into the summary without much scepticism. I now cross-check any X-only claim before it leaves my notes.

On the consumer app itself, the workflow held up under sustained use. I ran a week of research tasks through the web interface, and the answer format is more disciplined than it was a year ago: claims come with links, and follow-up questions keep thread context without drifting. Memory across chats is on by default and I found it remembered preferences I set in week one, which is either a feature or a surprise depending on your appetite for persistent assistants. I also tested file and PDF analysis on a 60-page contract PDF; extraction of clause text was accurate, though the summarisation flattened two conditions I had to go back and read myself. That matches my general experience: Grok is strong at retrieval and freshness, and merely adequate at careful analytical reading of dense documents.

Voice quality is closer to the leading voice products than it was in the Grok 3 era. Speech-to-speech responses in my tests came back in roughly conversational time, and at $0.08 per minute the cost of a ten-minute hands-free session is 80 cents at list rates — cheap enough that I stopped thinking about it. Transcription at $0.10 per hour (REST) is among the lowest rates I have seen for a first-party speech-to-text endpoint.

On packaging, the $30 Premium tier is the number I point readers to: xAI says it bundles SuperGrok, Cursor Pro and X Premium+, previously $90 combined. That is a vendor claim about the old combined price, but the bundle itself is visible on the plan comparison, and the shared usage pool across Grok apps, the Grok tab and Cursor is a genuinely convenient billing simplification.

What it does not do

There is no self-hosted or on-premises deployment; everything runs on xAI infrastructure, and enterprise controls on the consumer plans are thin. The content policy is looser than OpenAI's or Anthropic's, which some individuals value and which I would not choose for customer-facing brand work without a review pass. X ownership cuts both ways: the data access is unmatched, but your prompts and usage sit inside that corporate structure, and plan terms have changed twice in twelve months. The API has no provisioned-throughput or reserved-capacity option, so latency at peak times is what it is; the Super tier advertises "priority model access at peak times", and the lower tiers do not. Multi-agent mode is slower than single-shot chat by a wide margin in my runs, and the image models, while fast, still trail dedicated art tools on prompt adherence for complex compositions. Grok Bot and Cursor sit inside the same subscription now, which means the bill for a heavy user is one number covering three products — convenient, but harder to attribute spend to a specific workflow than a plain API account.

Verdict

Taken as a whole, Grok earns its slot in my rotation rather than replacing anything. It is the assistant I reach for when freshness matters more than caution: breaking news, X-native topics, anything where a training cutoff is disqualifying. The API is priced competitively at the low end, the long-context band punishes sloppy chunking, and the new bundle materially improves the value case for anyone already paying for Cursor or X. I rate it 4.1, weighted around strong core functionality across chat, search and multimodal generation, good value at the $8 and $30 tiers, a polished cross-platform experience, an API ecosystem that is growing quickly through the Cursor bundle, and acceptable but unexceptional reliability signals given the pace of plan and model churn.

Sources

• https://x.ai/grok
• https://x.ai/pricing
• https://x.ai/news/x-subscription
• https://docs.x.ai/developers/pricing
• https://docs.x.ai/developers/models
• https://docs.x.ai/developers/grok-4-5
• https://docs.x.ai/developers/tools/web-search
• https://justinmckelvey.com/blog/supergrok-pricing
• https://grabaicredits.com/en/pricing/grok-pricing
• https://felloai.com/ko/grok-pricing
• https://cabina.ai/blog/grok-api-pricing
• https://therouter.ai/blog/xai-grok-api-integration-guide