The infrastructure xAI has actually built
xAI builds Grok, and its developer documentation layer registers pricing specifications across five product surfaces: the text models, the Grok Build coding agent, the Imagine subsystem for image and video, Voice for speech, and the server-side tools. What cannot be extracted from this record is any claim about Grok inside the X app, and the reason is documented below.
- text, image, video and voice pricing in one record
- a price per thousand tool calls
- no geographic-coverage specification
Last checked: This is a reference page. It is re-checked against the vendor sources and updated when a new version ships. Change log
Model families
One text model line, whose newest member carries the smallest operating window.
Versions
| Versions | Deployed | Status | Model families |
|---|---|---|---|
| Grok 4.3 | current | Grok 4 | |
| Grok 4.5 | current | Grok 4 | |
| Grok 4.6 | current | Grok 4 | |
| grok-build-0.1 | current | Grok 4 |
The segment this server could not extract, and why we document it
x.ai returns a 403 status to our server on every path, both through WebFetch and through curl with a browser user agent. The robots.txt on that same domain permits crawling and even names several AI crawlers explicitly as permitted. So this is not a rule we would be circumventing, it is a bot-protection layer that has caught this server as well.
Everything in this record is therefore extracted from docs.x.ai, which is fully open, and no figure about a consumer Grok subscription, Grok inside the X app, or grok.com is logged here. If you encounter those figures elsewhere, request their source. We are in the identical position on the OpenAI record and document it there as well.
The most complete pricing table in this group
On one record, xAI logs pricing specifications for text, image, video, text-to-speech, speech-to-text, and live voice side by side. But the genuinely rare segment sits lower: a price for each server-side tool invocation. Web search, X search, and code execution run $5 per thousand calls, searching attached files runs $10, and searching document collections runs $2.50.
This matters to anyone building an agent, because in an agentic pipeline the model itself decides how many times to invoke a tool, and that segment of the invoice is normally undocumented anywhere and produces a surprise at the end of the billing period. No other vendor in this reference publishes these figures. Add a 2x priority-processing premium and a 20 percent batch-execution discount, both documented as numbers.
Where that same transparency turns on itself
The same batch-discount table shows Grok 4.5 excluded from the discount list, and the Batch API record states it more bluntly: this model is unsupported and its requests are rejected. The vendor own flagship is the one unit with no low-cost, high-volume lane.
The second case surfaces from the model specification data on the same catalogue record: the grok-latest alias points to Grok 4.3, not 4.5. For a vendor documenting everything else this transparently, both cases are unexpected, and both are logged in their correct location, just not where a buyer typically looks.
Documented strengths
- The most complete published pricing table in this group: text, image, video, speech and tools in one record
- The only vendor in this reference documenting a cost per thousand server-side tool calls
- Grok Build, an official command line agent with plugins and connections to external tools
- OpenAI-compatible interface tier, so code migration cost runs close to zero
- A 20 percent batch discount and 2x priority processing, both published as numbers
Known weaknesses and limits
- x.ai returns a 403 status to this server on every path, so this reference logs no consumer-facing data about Grok.
- No geographic-coverage specification exists in the open documentation, so the Iran column stays unchecked.
- Batch execution rejects the vendor own flagship while cheaper models receive the discount.
- No maximum output-token specification is published for any model.
- The grok-latest alias points to 4.3 rather than the flagship, which is easily mistaken.
- Its image, video, and voice models carry no verified row on any leaderboard we can read, so they are absent from our rankings.
Technical verdict
If you are building agents and want to model cost in advance, this is the clearest pricing table logged in this reference, and that alone is sufficient reason to take xAI seriously. But until the consumer surface of this vendor becomes readable to this server, everything documented here concerns the API layer, not what an ordinary user purchases.
Frequently raised questions, with documented answers
What is the pricing specification for Grok
Grok 4.5 runs $2 input and $6 output per million tokens, and $4 and $12 from 200,000 tokens upward. Grok 4.3 runs $1.25 and $2.50. We hold no consumer subscription price, because x.ai returns 403 to this server.
What is Grok Build
The xAI command line coding agent, running on Grok 4.5 itself, with an interactive interface, a headless mode for scripts, and connections to external tools through an agent protocol.
Does Grok operate from Iran
Unchecked, no claim on record. The developer documentation carries no geographic-coverage specification and x.ai own pages do not open for this server, so any firm answer today has no source behind it.