Grok Usage Limits (2026): x.ai Publishes No Number for Any Plan, Only a $0, $30 and $100 Ladder

The short answer

Grok's usage limits are not published. As of September 2026, x.ai prices three consumer plans and describes what each one allows in adjectives: Free at $0 has "generous limits", SuperGrok at $30 a month has "higher rate limits across all features", and SuperGrok Plus at $100 a month has "significantly higher usage across Chat, Imagine, Voice & Build". There is no message count, no window length and no reset period anywhere on the pricing page.

That is the finding, and it is worth more than a made-up number. The one place x.ai does print figures is the developer platform, where the flagship model is $2.00 per million input tokens. Everything below comes from x.ai's own pages.

What x.ai actually publishes

Here is the consumer ladder, quoted from the plan cards and the comparison table on x.ai/pricing:

That last block is the part people miss. x.ai names seven tiers on its own pricing page and puts a price on three of them. SuperGrok Lite, SuperGrok Heavy and Business appear as columns in the comparison table with feature checkmarks and no number above them, which means you cannot build a full ladder from this page no matter how carefully you read it.

One honesty note on sourcing. Everything here was read directly from x.ai in a browser, because the site blocks plain automated fetches. The consumer app at grok.com has its own help pages that I could not read the same way, so if a per-window prompt figure is published anywhere, it is not on the pricing page. Figures like that circulate widely in third-party write-ups, including in an older post of mine, and none of them appear on x.ai's pricing page as of this read.

How the tiers actually differ

Every tier is described relative to the tier below it, and the bottom rung is described relative to nothing. Free gets "generous limits". SuperGrok gets "higher rate limits". SuperGrok Plus gets "significantly higher usage". There is no base quantity, and there is not even a multiplier, so there is no arithmetic that recovers a number. Compare that to a ladder built on multipliers, where a tier described as three times the base at three times the price at least tells you the price per unit is flat. Here, "higher" tells you the direction of the change and nothing about its size, which means you cannot size a plan before you buy it and you cannot tell after a quiet month whether you overpaid.

What the comparison table does separate cleanly is features, not quantities. It splits the tiers on models, on whether you get Grok Build and Grok Bot, on whether real-time web and X search is full or "Limited", on connectors, on team and admin controls like seat management, consolidated billing and role-based access, on security terms including no training and custom data retention, and on support. So feature access you can size in advance with real confidence. Capacity is the single axis you have to discover by using it, and the free tier is the cheapest place to discover it.

The one surface where x.ai prints numbers

The developer platform is a different world. Read on September 21, 2026, x.ai/api publishes a full rate card, serves from https://api.x.ai/v1, and accepts existing OpenAI or Anthropic client libraries, so switching a test workload over is a base URL and a key rather than a rewrite.

• grok-4.7, the flagship as of that date: 500k context, $2.00 per million input tokens and $6.00 per million output tokens. grok-4.6 and grok-4.5 are still served for text, reasoning and coding.
• Imagine: grok-imagine-image-2.0 from $0.04 an image at 1K to 2K, grok-imagine-video-1.5 from $0.08 a second at 480p to 1080p, and the legacy grok-imagine-image at $0.02 an image.
• Voice: $0.08 a minute speech-to-speech, $15.00 per million characters text-to-speech, and $0.10 an hour speech-to-text over REST ($0.20 streaming).

Enterprise terms on the developer side add SSO, audit logging, SOC 2 Type II, HIPAA eligibility with a BAA, data residency, invoicing, zero data retention, and, notably, custom rate limits. The company will negotiate a number with you. It just will not print one on the consumer page. Grok is also resold through Azure AI Foundry, Oracle OCI Generative AI and Google Vertex Model Garden, which matters if your compliance review is already done for one of those. If you are pricing a workload rather than a seat, the same per-million comparison for the other vendors is in OpenAI API pricing.

What to do about an unpublished cap

Buy the cheapest tier that has the features you need, use it hard for a month, and let your own usage pick the tier. That is a worse experience than reading a table, and it is the actual state of the product as of September 2026.

Two specifics make that month cleaner. First, start on Free rather than on $30, because the free tier is a real product and the gap you are testing is capacity, not capability. What free visibly does not give you is full real-time web and X search, and that is the thing most people are actually buying Grok for, so notice whether the "Limited" version is the thing that runs out first. I went through what $0 covers in detail in Is Grok free. Second, decide up front whether you are buying volume or speed. SuperGrok Plus at $100 sells both, and only one of them is measurable in a week.

The $30 price is the part worth sitting with. Most of the assistants a small business compares Grok against sit at $20 a month, so SuperGrok is a 50% premium on a plan whose ceiling nobody will tell you. That premium is defensible if the live X feed is the reason you are there and irrelevant if it is not, which is the argument I make in Is Grok worth it.

Grok is not the outlier any more

Anthropic stopped publishing a message count some time ago and now describes allowances as multiples of a free tier it also does not quantify, which I wrote up in Claude usage limits. Google publishes multipliers over a "standard" it never defines. Cursor publishes multipliers of a base it never states. x.ai has gone one step further and dropped the multipliers too, which is the most honest version of the same move: at least nobody is pretending the number is recoverable.

The practical effect across the whole category is that limits have moved from something you plan against to something you discover. If you are budgeting a stack across several of these tools at once, the side-by-side of where every vendor's caps land, and which cells say "not published", is in AI usage limits, which gets re-read against the vendor pages every Monday for exactly this reason.

Last verified: September 21, 2026, against x.ai/pricing and x.ai/api, both read directly. Prices and published limits on this page are as of September 2026 and x.ai is the only source used for them.