Applied artificial intelligence, neural systems, and smart agents.
Fusing generative AI, multi-agent frameworks, and high-throughput data processing to automate business workflows.
Grok 4.5
Grok 4.5 records a rate of $2 per million input tokens and $6 per million output tokens, carries a 500,000 token context window, and sits twelfth of 113 models on the WebDev Arena leaderboard at a score of 1554. One structural condition on this pricing never surfaces in comparison tables: the moment a request prompt reaches 200,000 tokens, every token in that request is recalculated at double the rate.
- 1554 on WebDev Arena
- $6 output token rate
- 500k token window capacity
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 threshold one token triggers
The xAI pricing page defines two columns for every model: short context and long context. Grok 4.5 is specified at $2 and $6 in the first column and $4 and $12 in the second. The sentence recorded beneath the table carries more weight than the table itself: models with long context pricing bill the long context rate across every token in a request once its prompt crosses the threshold.
So this is not a gradual step, it is a cliff edge. A 199,999 token prompt draws an input cost of $0.40; a 200,000 token prompt draws $0.80. One additional token, double the bill, and the output on that same request jumps from $6 to $12. If a workload oscillates near that boundary, real cost swings between the two rates and averaging across them yields no reliable figure.
Reasoning cannot be stripped from the architecture
The xAI reasoning guide lists three levels for the reasoning_effort parameter on Grok 4.5, sets the default to high, and records an explicit statement: reasoning cannot be disabled. Reasoning tokens are billed exactly like any other output token.
Set that against the pricing structure. For heavy workloads this is a useful property. But if the workload consists of many short, repeated answers, the model spends output tokens on reasoning before every response and there is no way to halt that process. Grok 4.3 accepts the same parameter with a value of none — meaning the cheaper model in the same family builds a reasoning-disable path into its architecture and the flagship does not.
The flagship the batch execution lane rejects
xAI applies a 20 percent discount to the batch execution lane for Grok 4.3 and the three 4.20-family models. The Batch API documentation carries a separate warning for Grok 4.5: this model is not supported on that lane and its requests are rejected. So the one model in the catalogue best suited to high-volume workloads is precisely the one without a cheap high-volume lane.
This is the exact pattern observed on Fable 5, where the more expensive model in the family was left off the vendor own batch capability list. Notably, both vendors recorded this constraint in their technical documentation and neither repeated it on any sales page.
Its position on this reference leaderboard
In the coding category it scores 10.8 out of 100, the seventh and last ranked row in this table. Almost the entire score is supplied by one column: tooling maturity, at 10 of 15 weight. It scores zero on three other columns, because it sets the floor on all three: the web-dev leaderboard, the general text leaderboard, and context window capacity.
And a note on our own method, which we would rather record explicitly. The gap between Grok and DeepSeek V4 Flash on the web-dev leaderboard is 31 points, 1554 against 1585. DeepSeek draws 9 points from that forty-weight column and Grok draws zero. Min-max normalization always assigns zero to the floor of a table, regardless of how small the gap to the row above it is. So the 10.8 score should be read as last row in a table of seven high-tier models, not as weak performance.
And an apparent inconsistency that is not one: it holds twelfth position of 113 on the web-dev table and thirty-fourth of 389 on the general text table, where xAI own older grok-4.20-beta1 sits above it at 1475. The vendor own documentation confirms this same structure: Grok 4.5 is introduced as our flagship model for code and everything else. Its architecture is optimized for code, and in general conversation it has not surpassed its own predecessor.
Specifications
Every number here comes from the vendor's own page, with that page linked beside it.
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: No page carrying a country-access list for this service could be reached: the x.ai endpoint returns a 403 status to our measurement server on every request path, while its robots.txt permits crawling of that same domain. The documentation subdomain, docs.x.ai, does respond but carries no country list at all. So this path remains recorded as unmeasured.
Documented strengths
- A $6 output rate against $25 for Opus 5, with a 138-point gap on the WebDev leaderboard
- A rating-per-output-dollar figure of 259, roughly 3.8 times Opus 5
- The official Grok Build command-line agent, which runs directly on this same model
- Native image input support and an interface compatible with the OpenAI specification
- A cached input rate of $0.30 for prompts built on a fixed prefix
Room for improvement and technical considerations
- A 500,000 token window capacity, the smallest figure in this reference coding table and half of Grok 4.3, which also carries a lower cost.
- Past the 200,000 token threshold, the entire request is recalculated at double the rate, not just the tokens above the threshold.
- Reasoning cannot be stripped from the architecture and defaults to high, so every response consumes extra output tokens.
- The batch execution lane rejects this model, while cheaper models from the same vendor draw a 20 percent discount.
- No figure for maximum output tokens is published in the documentation.
- The Iran reachability column is recorded empty, because the x.ai endpoint returns a 403 status to our measurement server and docs.x.ai carries no country list.
Technical verdict
For a coding workload where prompts stay under the 200,000 token threshold, this model is the cheapest route to a high tier with real tooling built around it, and that conclusion holds seriously. But for a workload that feeds a full code repository into the model, our recommendation sits inside xAI own catalogue rather than outside it: Grok 4.3 carries double the window capacity, half the output rate, and the batch execution discount. That is an unusual conclusion to record about a flagship model, but it is what the figures show.
Frequently raised questions, with documented answers
What is the pricing rate for Grok 4.5
Two dollars per million input tokens and six dollars per million output tokens, as long as the prompt stays under the 200,000 token threshold. Above that, the rate shifts to four and twelve dollars, and it applies to every token in that request.
What happens if a prompt reaches exactly 200,000 tokens
The entire request is recalculated at the long context rate. A 199,999 token prompt draws an input cost of $0.40, and a prompt at exactly 200,000 tokens draws $0.80. A difference of one token, a doubled bill.
Is it superior to Opus 5 for coding
Not on the WebDev leaderboard: 1554 against 1692. But its output rate is a quarter of Opus 5, so at volume and without a requirement for top-tier quality, it represents a reasonable trade-off. In this reference table it holds seventh position and Opus 5 holds first.
Is Grok reachable from Iran
Unmeasured. The x.ai endpoint returns a 403 status to our measurement server on every request path, and its developer documentation carries no country list, so its status cannot be recorded as either open or blocked.