xAI launched Grok 4.5 on July 8, announcing the model on its newsroom page alongside a cluster of coding and agentic benchmark results, positioning it squarely at software engineering and technical document generation workloads xAI.

The model was trained in conjunction with Cursor, on datasets spanning coding, science, engineering, and math, according to xAI xAI. Training ran across tens of thousands of NVIDIA GB300 GPUs, with data pipelines built around deduplication, quality scoring, and domain-focused curation. Reinforcement learning for the model covered hundreds of thousands of tasks centered on multi-step software engineering and other technical work, xAI said.

On benchmarks, Grok 4.5 posted a #1 ranking on Harvey's Legal Agent Benchmark, a domain-specific evaluation used by legal-tech vendor Harvey to assess agentic reasoning on legal workflows. On coding-specific evaluations, xAI reported 62.0% on DeepSWE 1.0 within its own test harness, and 53% on DeepSWE 1.1 when run independently by DataCurve using the mini-swe-agent harness — a gap that reflects the difference between vendor-run and third-party-run evaluations on a newer benchmark revision. The model also scored 83.3% on Terminal Bench 2.1, a benchmark measuring agentic command-line task completion, and achieved a 64.7% resolve rate on SWE Bench Pro, a harder variant of the widely used SWE Bench suite that evaluates whether a model can resolve real-world GitHub issues end to end.

Token efficiency featured prominently in xAI's release. Grok 4.5 averages 15,954 output tokens per SWE Bench Pro task, which xAI describes as roughly 4.2 times fewer than Anthropic's Opus 4.8 (max), which averages 67,020 tokens on the same benchmark. The model is served at 80 tokens per second.

Grok 4.5 is now the default model in Grok Build, xAI's application-generation product. Within Grok Build, the model can construct complex Excel models that incorporate web research, multi-sheet formula chains, and embedded sticky notes intended to preserve context for future edits. It can also generate PowerPoint content using native shapes to construct diagrams and slide layouts, rather than relying on templated placeholders or externally rendered images.

The token-efficiency framing is worth sitting with for a moment. Inference cost and latency have become as central to model selection as raw benchmark accuracy, particularly for agentic workloads that chain dozens or hundreds of tool calls in sequence — a 4x reduction in output tokens per task, if it holds across production workloads rather than just the cited benchmark, changes the unit economics of running these agents at scale far more than another few points on a leaderboard would.

The discrepancy between the 62.0% DeepSWE 1.0 score generated in-house and the 53% DeepSWE 1.1 score generated externally is a reminder, familiar to anyone who has tracked benchmark reporting across several model generations now, that vendor-reported numbers and independently reproduced numbers routinely diverge, sometimes because of harness differences and sometimes because of benchmark version changes themselves. Neither number is necessarily wrong. They measure somewhat different things, and readers evaluating Grok 4.5 against competitors should weight the DataCurve figure more heavily precisely because it was not run by xAI itself.

The emphasis on Excel and PowerPoint generation signals where xAI expects near-term enterprise adoption to concentrate. Coding agents have dominated the last two years of model releases, but office-document generation — spreadsheets with live formula logic, slide decks with genuine diagrammatic structure rather than boilerplate — touches a much larger population of knowledge workers who will never open a terminal. Whether Grok Build's Excel and PowerPoint capabilities hold up against the fiddly, formula-dependent reality of enterprise finance and consulting workflows is the kind of claim that tends to look better in a demo than in a live model with three years of legacy formatting baked in, and it is the sort of thing best judged by hands-on use rather than a press release.

The Harvey benchmark placement is notable less for the raw score than for what it signals about vertical-specific competition. Legal, like coding, has become a proving ground where general-purpose frontier models are increasingly measured against domain benchmarks built by specialized vendors rather than generic academic test sets, and a #1 ranking there gives xAI a concrete talking point with enterprise legal buyers that a SWE Bench score alone would not.