xAI has launched Grok 4.5, a new general-purpose model aimed at software engineering, long-running agent tasks and professional knowledge work. The model is available through the xAI API and console, is now the default model in Grok Build, and is also offered through Cursor. xAI says access in Grok Build and Cursor will be free for a limited period, although it has not specified when that introductory offer will end.
The release gives developers another frontier model to evaluate for coding agents and tool-using workflows. xAI has set API pricing at US$2 per million input tokens and US$6 per million output tokens. It reports serving speeds of about 80 tokens per second, but real application latency will also depend on prompt size, tool calls, network conditions and the complexity of each task.
Engineering and agent capabilities
xAI describes Grok 4.5 as its strongest model for coding, science, engineering and mathematics. The launch material highlights work across Rust, C and C++, along with the ability to build complete applications from a single prompt. It also shows the model working with spreadsheets, presentations and documents, extending the release beyond conventional code generation.
The company published results across DeepSWE, SWE Marathon, Terminal Bench and SWE Bench Pro. Grok 4.5 led some comparisons and trailed other models in others. Those results should be read carefully: benchmark harnesses, reasoning settings and token budgets can materially affect outcomes, and xAI notes that some competitor figures came from other developers' system cards or public leaderboards.
For buyers, the practical question is whether Grok 4.5 can complete real repository work reliably, not whether it wins every benchmark. Evaluation should include regression risk, tool-call accuracy, review time, security boundaries and the cost of failed or repeated runs.
Training and efficiency claims
xAI says Grok 4.5 was trained across tens of thousands of NVIDIA GB300 GPUs. Its training process combined filtered and deduplicated data with reinforcement learning across hundreds of thousands of multi-step engineering and technical tasks. The company says its asynchronous training system can support agent rollouts lasting many hours.
Efficiency is a central part of the pitch. xAI reports that Grok 4.5 used about 15,954 output tokens per SWE Bench Pro task in one comparison, roughly 4.2 times fewer than the cited Opus 4.8 result. It also claims about twice the token efficiency of comparable leading models more broadly. These are vendor-reported figures and may not translate directly to every codebase, agent framework or production workload.
What developers should test
Teams considering Grok 4.5 should run a controlled evaluation using their own repositories, tool permissions and quality gates. Useful measures include successful task completion, human review effort, time to resolution, output-token consumption and the frequency of unsafe or unnecessary actions. The advertised input and output prices make the model straightforward to compare with alternatives, but total workflow cost can rise when an agent uses many tools or retries.
The release expands xAI's position from consumer chat into coding and enterprise-style work. Availability across xAI's own tools, a direct API and Cursor gives the company several routes to adoption. The unanswered questions are how consistently the model performs outside launch benchmarks and how its reliability changes on longer, less structured tasks.