The Grok 4 family
For a long-input workload, Grok 4.3 belongs in deployment, not 4.5. The newer version in this family carries a 500,000-token window against Grok 4.3 one million, its output rate is $6 against $2.50, and batch execution is not supported on it at all.
- 4.5 window is half of 4.3
- output rate $6 against $2.50
- batch discount active only on older versions
Last checked: This is a reference page. It is re-checked against the vendor sources and updated when a new version ships. Change log
The timeline, and an upgrade that revokes three parameters
The current xAI catalogue prices these versions: Grok 4.5 with a 500,000-token window, Grok 4.3 with one million, and three 4.20-generation models that also carry a one-million-token window, one of them multi-agent. Migrating from 4.3 to 4.5 revokes three parameters from your architecture and grants one.
It revokes: half the context window, the 20 percent batch-execution discount active on 4.3 and disabled on 4.5, and the ability to disable reasoning. The reasoning guide states explicitly that this parameter cannot be disabled on 4.5, while 4.3 accepts the same parameter with a value of none. It grants: a substantially higher benchmark score on the coding workload, 1554 against 1355 on the WebDev Arena leaderboard, and a wider published rate-limit ceiling.
The rate rose as well, not fell: input from $1.25 to $2, output from $2.50 to $6. We logged the same direction in the GLM family and the reverse in the Opus family. So "newer version means lower rate" is not a fixed rule in this market, however often summaries frame it that way.
The grok-latest identifier does not resolve to 4.5
The documentation states that an unsuffixed identifier resolves to the latest stable version of that model line. In the model data the same catalogue page publishes, grok-latest is logged among the aliases of Grok 4.3, not 4.5; the aliases for 4.5 are grok-4.5-latest and grok-build-latest.
The operational consequence: any architecture that has grok-latest hardcoded and assumes it is deployed on the flagship is actually running on 4.3. Its invoice is lower, and so is its coding benchmark score. This is stated on no product page and can only be extracted by placing two rows of data side by side.
Which version belongs in deployment
For a coding agent with short requests, 4.5. For any workload with long inputs or overnight batch execution, 4.3: double the context window, an output rate under half the price, and a 20 percent batch discount on top. For a workload that requires no reasoning at all and only needs a fast response, the 4.20 generation even ships a separate non-reasoning model under its own identifier.
One constraint applies to all three groups: none escapes the 200,000-token threshold. Across every version in this family, any request whose length reaches that number bills every token in that request at double the rate.
Release timeline
Each version with its own release date, and what changed against the one before it.
- Grok 4.3 current
- Grok 4.5 current Half the context window capacity of Grok 4.3, at more than double its output rate.
- Grok 4.6 current
- grok-build-0.1 current
Using it from Iran
This column carries a date and says how it was checked, because most listicles guess it. Where we have only read a vendor policy page, the note below says exactly that.
- Reachable
- not checked
- Payment
- not checked
- Free tier
- no
How we checked: x.ai returns a 403 response to our server on every path, so a page carrying a country list has not been accessible to us. The developer documentation is open but publishes no such list.
Documented strengths
- Two distinct deployment paths in one family: a higher-rate coding model and a lower-rate long-input model
- Grok 4.3, with a one-million-token window and a $2.50 output rate, the lowest-cost entry point into this family
- A 20 percent batch-execution discount on 4.3 and on all three 4.20-generation models
- The 4.20 generation ships both a non-reasoning model and a multi-agent model under their own identifiers
Known weaknesses and limits
- The newest version in this family carries the smallest context window: 500,000 tokens against one million.
- The rate rose along the production line: the output rate on 4.5 is more than double 4.3.
- Batch execution rejects Grok 4.5, so no volume discount is available on the flagship.
- The 200,000-token threshold doubles the entire request on every version, not only the tokens above the threshold.
- No maximum output-token figure is published for any version.
- Access from Iran is unverified for every version, because x.ai returns a 403 response to our server.
Technical verdict
This is the only family in this reference where the flagship is the wrong deployment choice for most workloads. If your input is a full repository or a long document, 4.3 belongs in deployment; deploy the flagship only when the workload is coding and request length stays short.
Frequently raised questions, with documented answers
Does Grok 4.5 outperform Grok 4.3
For coding, yes, 1554 against 1355 on the web-dev leaderboard. For any other workload it depends: 4.3 carries double the context window, an output rate under half the price, and an active batch discount.
Which model does grok-latest resolve to
In the model data published by the xAI catalogue page, grok-latest is logged among the aliases of Grok 4.3. Deploying the flagship requires naming its identifier explicitly in the configuration.
Which version in this family is the lowest cost
Grok 4.3, at $1.25 input and $2.50 output, with a further 20 percent reduction when the workload runs as a batch. The cheaper catalogue model, Grok Build 0.1, is not part of this family and carries a 256,000-token window.