Grok 4.5
Grok 4.5 charges $2 per million input tokens and $6 per million output, carries a 500,000 token window, and sits twelfth of 113 models on WebDev Arena at 1554. One condition on that price never makes it into the comparison tables: the moment a request prompt reaches 200,000 tokens, every token in that request is billed at double.
- 1554 on WebDev Arena
- $6 per million output tokens
- 500k token window
Last checked: This is a reference page. It is re-checked against the vendor sources and updated when a new version ships. What changed
The step that one token triggers
The xAI pricing page gives every model two columns, short context and long context. Grok 4.5 is $2 and $6 in the first and $4 and $12 in the second. The sentence under the table matters more than the table: models with long context pricing bill the long context rates for all tokens in a request once its prompt reaches the threshold.
So it is not a step, it is a cliff. A 199,999 token prompt costs $0.40 of input; a 200,000 token prompt costs $0.80. One extra token, twice the bill, and the output on that same request moves from $6 to $12. If your workload sits near that line, your real cost jumps between two rates and an average across them tells you nothing.
Reasoning cannot be turned off
The xAI reasoning guide lists three levels of reasoning_effort for Grok 4.5, sets the default to high, and states plainly that reasoning cannot be disabled. Reasoning tokens bill exactly like any other output token.
Put that next to the price. For heavy work it is a feature. But if your workload is many short answers, the model spends output tokens thinking before each one and you cannot stop it. Grok 4.3 accepts the same parameter with a value of none. The cheaper model in the family lets you switch reasoning off and the flagship does not.
The flagship the Batch API rejects
xAI gives Grok 4.3 and the three 4.20 models a 20 percent batch discount. The Batch API documentation carries a separate warning for Grok 4.5: the model is not supported and requests are rejected. The one model in the catalogue you would want for volume work is the one with no cheap volume lane.
This is the same shape we found on Fable 5, where the more expensive model was left off its vendor batch feature list. Both vendors wrote it down in their documentation and neither repeated it anywhere a buyer would look.
Where it lands in our table
It scores 10.8 out of 100 in the coding category, seventh and last of the ranked rows. Almost all of that comes from one column: tooling maturity, 10 of 15. It takes zero in three others, because it is the floor of all three: the web-dev leaderboard, the general text leaderboard and the context window.
And one point against our own method, which we would rather write than have somebody find. Grok sits 31 points behind DeepSeek V4 Flash on the web-dev leaderboard, 1554 against 1585. DeepSeek takes 9 points from that forty-weight column and Grok takes zero. Min-max normalisation gives the floor of a table nothing, however close the floor is to the row above it. So read 10.8 as "last in a table of seven high-tier models", not as "weak".
And an apparent contradiction that is not one. It is twelfth of 113 on WebDev and thirty-fourth of 389 on the general text board, where xAI own older grok-4.20-beta1 sits above it at 1475. The vendor documentation says the same thing in its own words: Grok 4.5 is introduced as "our flagship model for code and everything else". It was built for code, and in general conversation it has not beaten its own predecessor.
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: We could not read a page that would carry a country list: x.ai returns 403 to our server on every path while the same domain robots.txt allows crawling. docs.x.ai does answer and carries no country list at all. So this stays unchecked.
What it is good at
- $6 output against $25 for Opus 5, with a 138 point gap on the WebDev leaderboard
- 259 on rating per output dollar, roughly 3.8 times Opus 5
- Grok Build, an official command line agent running on this same model
- Native image input and an OpenAI-compatible interface
- A cached input rate of $0.30 for prompts built on a fixed prefix
Where it falls short
- The 500,000 token window is the smallest in our coding table and half of Grok 4.3, which also costs less.
- Past 200,000 prompt tokens the entire request bills at double, not just the tokens above the line.
- Reasoning cannot be disabled and defaults to high, so every answer carries extra output tokens.
- The Batch API rejects this model while cheaper models from the same vendor take a 20 percent discount.
- No maximum output token figure is published anywhere in the documentation.
- The Iran column is empty because x.ai returns 403 to our server and docs.x.ai carries no country list.
Our take
If you write code and your prompts stay under 200,000 tokens, this is the cheapest route to a high-tier model with real tooling around it, and we take it seriously. If you feed whole repositories to a model, our recommendation sits inside xAI own catalogue rather than outside it: Grok 4.3 has twice the window, half the output price and a batch discount. That is a strange sentence to write about a flagship, but it is what the numbers say.
Questions people actually ask
What does Grok 4.5 cost
Two dollars per million input tokens and six per million output, as long as the prompt stays under 200,000 tokens. Above that it is four and twelve, and the higher rate applies to every token in that request.
What happens if my prompt hits exactly 200,000 tokens
The whole request bills at the long context rate. A 199,999 token prompt takes $0.40 of input and a 200,000 token prompt takes $0.80. One token of difference, twice the invoice.
Is it better than Opus 5 for coding
Not on the WebDev leaderboard: 1554 against 1692. But its output costs a quarter of Opus 5, so at volume and without a top-tier quality requirement it is a reasonable trade. In our table it is seventh and Opus 5 is first.
Does Grok work from Iran
Unchecked. x.ai returns 403 to our server on every path and its developer documentation carries no country list, so we can write neither open nor blocked.
Sources
- xAI models vendor source 12 August 2026
- xAI pricing vendor source 12 August 2026
- xAI Batch API vendor source 12 August 2026
- xAI reasoning guide vendor source 12 August 2026
- Grok Build vendor source 12 August 2026