xAI released Grok 4.5 on July 8, positioning the model directly at software engineering and technical work. The company trained it alongside Cursor, an AI code editor, on datasets spanning coding, science, engineering, and math. Training used tens of thousands of NVIDIA GB300 GPUs — specialized processors for AI — with careful data filtering and domain-focused curation throughout xAI.

To refine how the model handles real-world technical tasks, xAI used reinforcement learning — a technique that teaches AI by rewarding correct actions — on hundreds of thousands of problems centered on multi-step coding and engineering work.

Benchmark Performance

Grok 4.5 claimed top ranking on Harvey's Legal Agent Benchmark, a test used by legal-tech vendor Harvey to evaluate whether AI can handle legal workflows. On coding tests, the numbers tell a messier story. xAI reported 62.0% on DeepSWE 1.0 when running the test itself, but when an independent lab (DataCurve) ran the same model on a newer version of that benchmark (DeepSWE 1.1), it scored 53%. The model also achieved 83.3% on Terminal Bench 2.1 — which measures whether AI can complete command-line tasks — and a 64.7% resolve rate on SWE Bench Pro, a harder variant that tests whether the model can fix real bugs in GitHub repositories.

Vendor-reported scores and independently tested scores routinely diverge, sometimes because of how the test is set up and sometimes because of benchmark version changes. Neither is necessarily wrong; they measure somewhat different things. If you are evaluating Grok 4.5 against competitors, the DataCurve figure carries more weight precisely because xAI did not run it themselves.

Efficiency: The Real Story

Token efficiency featured prominently in xAI's announcement, and for good reason. Tokens are chunks of text or code — think of them as the basic units the model processes. On SWE Bench Pro, Grok 4.5 averages about 15,954 tokens to complete a task. By contrast, Anthropic's Opus 4.8 averages around 67,020 tokens for the same work. That is roughly 4.2 times fewer tokens — and far fewer tokens means lower costs and faster responses when running these models at scale.

The model serves responses at 80 tokens per second, making it practical for real-time use.

This efficiency gap matters more than it might sound. In the last few years, as AI has moved from simple chat to handling complex, multi-step tasks — stringing together dozens of tool calls, running queries, iterating on results — the cost and speed of running those chains has become as important as raw accuracy. If Grok 4.5 truly uses 4.2 times fewer tokens in production (not just on benchmarks), it fundamentally changes the economics of deploying these agents at scale.

Where xAI Sees Adoption

Grok 4.5 is now the default model in Grok Build, xAI's application-generation product. The model can construct complex Excel spreadsheets with live formulas, multi-sheet logic, and context notes that preserve information for future edits. It can also generate PowerPoint slides, creating actual shapes and diagrams rather than dropping in templates or flat images.

This focus on office tools signals where xAI believes near-term adoption will happen. Coding agents have dominated AI releases over the past two years, but spreadsheet and slide generation touch a far larger population of everyday knowledge workers — people who will never use a terminal or write code. Whether Grok Build can handle the messy reality of enterprise finance spreadsheets and ten-year-old formatting quirks is harder to judge from a press release. That kind of claim is best tested hands-on, in real workflows, with actual legacy data.

The legal benchmark ranking deserves attention for a different reason. Legal, like coding, has become a space where general-purpose AI models are evaluated against industry-specific tests built by specialized vendors rather than generic academic benchmarks. A top score there gives xAI a concrete selling point with enterprise legal teams that a generic coding benchmark alone would not provide.