xAI has launched Grok 4.5, its most capable model to date. The company says it was built specifically to handle coding, agentic tasks, and knowledge work, and was trained in partnership with Cursor, the popular AI-powered code editor.
The release puts xAI in direct competition with models like OpenAI’s GPT-4o and Anthropic’s Claude 3.5 Sonnet, which also target developers and enterprise users. What sets Grok 4.5 apart, according to xAI, is a combination of raw intelligence, token efficiency, and speed that together lower the cost of getting useful results.
The model is available now through Grok Build, Cursor on all plans, and the xAI API console. For a limited time, xAI is offering free usage in both Grok Build and Cursor.
What Grok 4.5 was built to do
xAI trained Grok 4.5 on datasets covering coding, science, engineering, and math. The goal was a model that could handle real engineering work rather than just benchmark tests. Key capabilities include:
- Complex coding tasks in Rust, C, and C++
- End-to-end app building from a single prompt
- Multi-step software engineering workflows
- Office work including Excel models, PowerPoint presentations, and Word documents
- Legal and knowledge work, where it scored first on Harvey’s Legal Agent Benchmark
On the office side, Grok 4.5 can build complex Excel models that pull live data from the web, use multi-sheet formulas, and even leave notes for future reference. In PowerPoint, it uses native shapes to build diagrams rather than relying on static images, which is a meaningful step up from what most AI tools currently produce.
How it was trained
Training ran across tens of thousands of NVIDIA GB300 GPUs. xAI says it put significant effort into data quality rather than just volume, using deduplication, quality scoring, and domain-focused selection to keep the training data both broad and high-signal.
The reinforcement learning setup covered hundreds of thousands of tasks focused on multi-step software engineering. Because the training stack is built for highly asynchronous runs, agentic rollouts can run for many hours while the model keeps learning in parallel. xAI says this approach produces more efficient reasoning on real engineering tasks rather than just faster pattern matching.
Speed and pricing
Grok 4.5 runs at 80 tokens per second. xAI also claims the model achieves roughly twice the token efficiency of comparable leading models, meaning it solves tasks in fewer steps. In practice, that combination of speed and efficiency is meant to reduce both wait times and costs.
Pricing is set at:
- $2 per million input tokens
- $6 per million output tokens
That positions Grok 4.5 as a competitive option for developers who run high volumes of requests. The doubled token efficiency claim, if it holds up in real-world use, would effectively halve the per-task cost compared to similar models at the same price point.
Where you can use it
Grok 4.5 is now the default model in Grok Build, xAI’s development environment. It’s also available in Cursor on all plans, which gives it immediate reach across a large developer user base. API access is available through the xAI console with a standard API key.
One notable exception: Grok 4.5 is not yet available in the EU across any xAI products or the API. EU availability is expected in mid-July, likely pending regulatory review under the EU AI Act and related compliance requirements.
Developers can get started at x.ai/cli, where free usage is available for a limited time.
