xAI's newly announced "Grok 4.5" is a model that pushes price to the forefront of the performance race. The announcement describes it as a top-tier model for coding, agentic work, and knowledge labor. API pricing is $2 per million input tokens and $6 per million output tokens. The more advanced models are used for long reasoning chains and code fixes, the larger a single response tends to grow. The real point of contention for Grok 4.5 isn't setting new top scores—it's how far the per-use cost of a frontier-class model can be pushed down.
API Pricing Starting at $2 per Million Tokens
Grok 4.5's API pricing is listed as $2 per million input tokens and $6 per million output tokens, consistent across both xAI's announcement page and its developer documentation. The model ID in the documentation is grok-4.5, with a context length of 500,000 tokens. xAI positions it as a flagship model for code, agentic tool-calling, and knowledge work.
This pricing matters most for tasks where output ballooning outpaces input. In agentic coding, the model writes a plan, reads files, drafts fixes, and rewrites again based on test results. Even when a user's instructions are short, the code, explanations, and additional reasoning the model returns can be long. A rate of $6 per million output tokens carries more weight for teams running large volumes of this kind of iterative work than it does for one-off chat pricing.
According to competitor pricing compiled by SiliconANGLE, Anthropic's Opus 4.7 and 4.8 cost $5 per million input tokens and $25 per million output tokens, Fable 5 costs $10 and $50, and OpenAI's GPT-5.6 Sol costs $5 and $30. Grok 4.5, at least by published-price comparisons, comes in well below these top-tier models' output rates. That said, OpenAI's lower-priced Luna model is listed at $1 and $6, so Grok 4.5's pricing alone doesn't redraw the entire market.
Cutting Output Tokens Matters More Than Speed
xAI states that Grok 4.5 runs at 80 TPS. Fast responses directly affect developer experience, but the token-efficiency claim carries more weight on the cost side. According to xAI, the average number of output tokens used to solve SWE Bench Pro tasks was 15,954, compared to 67,020 for the comparison model Opus 4.8 max. The announcement frames this as 4.2 times fewer output tokens.
If this gap holds up in real-world tasks, the price difference works twice over. First, the per-million-token output rate is lower. Second, fewer tokens are generated to complete the same job. In long refactoring efforts, investigating CI failures, or spec changes spanning multiple files, for example, the model repeatedly outputs explanations and code. A company's AI usage cost is determined not just by the per-token model price but by the total tokens consumed per task.
That said, xAI's figures are vendor-published numbers. In actual development environments, repository size, slow test suites, prompt design, and tool-integration failure rates all affect output volume. Grok 4.5's price gap is attractive, but adoption decisions require measuring whether the same savings materialize on your own tasks.
Benchmarks Speak to Competitive Range, Not Outright Wins
The benchmarks on the announcement page do not depict Grok 4.5 as "always the top model." On DeepSWE 1.0, as listed by xAI, Grok 4.5 scores 62.0%, below Fable max's 66.1% and GPT 5.5 xhigh's 64.31%. DeepSWE 1.1 came in at 53%. SWE Bench Pro sits at 64.7%, falling short of the Fable max and Opus 4.8 max figures shown in xAI's table.
On the other hand, Grok 4.5's SWE Marathon resolution rate is 29.0%, above Opus 4.8 max's 26.0% and Fable max's 24.0% as listed in the same table. Terminal Bench 2.1 comes in at 83.3%, close to Fable max's 84.3% and GPT 5.5 xhigh's 83.4%. In other words, Grok 4.5's pitch isn't "best across every benchmark." It's structured around delivering competitive-range performance at a lower API rate and with fewer output tokens.
Seen this way, it aligns with the announcement's emphasis on price. In the frontier model market, there are increasingly more scenarios where how many attempts you can make within the same budget matters more than a one-point benchmark difference. Code generation and business document creation aren't always tasks that get finished perfectly on the first response. The total cost—including fixes, verification, and reruns—becomes the number that drives model selection.
A Practical Sales Pitch Built on GB300 and Cursor
Regarding Grok 4.5's training, xAI mentions tens of thousands of NVIDIA GB300 GPUs, deduplication and quality scoring of data, domain-specific curation, and reinforcement learning at a scale of hundreds of thousands of tasks. It further describes a system that continues training while running hours-long agentic rollouts asynchronously. What the announcement emphasizes isn't short lab-style problems but practical tasks that involve long procedures.
The rollout points in the same direction. Grok 4.5 becomes the default model for Grok Build and is also available across all Cursor plans and in xAI's console. The announcement also touches on work examples in Word, PowerPoint, and Excel, treating coding tasks and office work as a continuous form of knowledge labor. Rather than leading with a standalone performance table, this is a sales approach that first secures the daily-use channels of developers and business users.
There are constraints too. As of the announcement, Grok 4.5 is not available in xAI products or the API console in the EU. xAI expects EU availability around mid-July 2026. Companies whose adoption timelines are affected by regulatory or regional rollout constraints can't decide on immediate adoption based on price alone.
What to Watch Next: EU Rollout and Real-World Billing
Grok 4.5's announcement pulls model pricing competition back from "monthly plan" thinking to "billing per task." xAI's consumer-facing pricing page shows SuperGrok at $30 per month, but the numbers developers and enterprises should be watching are different: $2 for input and $6 for output via the API, a 500,000-token context length, and how many tokens the model emits before completing a task.
What's troublesome for competitors is that Grok 4.5 isn't selling on top performance alone. It lands in the competitive range on benchmarks, then differentiates through per-token output pricing and generation volume. If this pattern holds, companies may find it easier to move away from designs that rely on expensive top-tier models at all times, toward designs that allocate models based on the balance of price and performance.
The numbers to watch next are clear: whether EU availability begins as planned; what actual billing settles at once free or discounted trials on Cursor or Grok Build end; and whether the token efficiency xAI has claimed holds up under external benchmarks and internal enterprise evaluations. Grok 4.5's pricing announcement is where the real, measured competition begins.