xAI announced Grok 4.5 on July 8, positioning it as a specialized tool for coding, spreadsheets, and technical document generation xAI.

The model was built alongside Cursor, a popular coding assistant, using training data focused on programming, science, engineering, and math. xAI ran the training across tens of thousands of NVIDIA processors and applied filtering techniques to ensure data quality. The team then used reinforcement learning — a method where the model learns from feedback on its own attempts — to train it on hundreds of thousands of tasks in software engineering and similar technical work xAI.

Grok 4.5 scored at the top of a legal industry benchmark designed to test how well AI assistants can handle law-related tasks. On coding benchmarks, it achieved 62.0% on DeepSWE 1.0 when tested by xAI, and 53% when tested independently by a third-party firm using a newer version of the benchmark. It also scored 83.3% on a test measuring how well the model can execute command-line tasks, and achieved a 64.7% success rate on resolving real GitHub issues — actual programming problems posted by developers online.

One number stands out in xAI's pitch: efficiency. Grok 4.5 uses roughly four times fewer output tokens — the units that determine how much computational work is needed — than Anthropic's competing Claude model on the same benchmark. In practical terms, this means running Grok 4.5 costs less per task and is faster to generate responses.

Within Grok Build, xAI's tool for creating applications, Grok 4.5 can now generate complex Excel spreadsheets with working formulas and research data built in. It can also create PowerPoint slides with actual diagram structures rather than just placeholder images.

The focus on efficiency matters because many of these AI models work by chaining together dozens or hundreds of small steps to solve complex problems — like looking up information, analyzing it, and writing a report. If a model can do the same work while using one-quarter the computational resources, companies running these AI systems at scale can lower their costs significantly. This advantage goes beyond what a slightly higher score on a performance test would suggest.

The gap between xAI's own test score (62.0%) and the independent test score (53%) is worth noting. When a company runs its own internal tests, it may have slight advantages in how those tests are structured or executed. A third-party test provides more neutral ground. Anyone comparing Grok 4.5 to other models should weight that independent 53% score more heavily, since it was not run by xAI itself.

The emphasis on Excel and PowerPoint signals where xAI expects businesses to adopt this technology. Coding assistants have gotten most of the attention over the past two years, but spreadsheet and slide generation could reach a much larger group of office workers. Whether these capabilities actually hold up when used with real company spreadsheets — complete with years of custom formatting, linked data, and hidden dependencies — remains a question best answered through actual hands-on testing rather than company demonstrations.

The top ranking on the legal industry benchmark is notable because it shows a pattern spreading across specialized fields. Coding was the first proving ground for these general AI models, measured against tests built by the coding community itself. Legal work is the next frontier, with tests built by legal-tech companies. A strong performance on such domain-specific benchmarks gives xAI a concrete selling point to businesses in those fields.