SAN FRANCISCO — SpaceXAI is keeping up a rapid release cadence for Grok Build, its terminal-based coding agent, pushing version 0.2.96 on July 10 with a batch of upgrades aimed at developers running long, autonomous coding sessions. The update lands with Grok 4.5, the company's strongest model to date, now serving as the default engine inside Grok Build.
What's New in v0.2.96
The release adds a new /goal slash command for setting a session objective, a refreshed /sessions flow, and a x.ai/pr/status check that reports whether an open pull request has reached the merge queue. System notifications now carry structured titles and bodies for cleaner rendering, a compact mode activates automatically on very small terminals, and the up arrow on an empty prompt browses prompt history that can be searched with /history.
Under the hood, SpaceXAI improved reliability of bash and file output, made subagent activity easier to follow with live task labels, and sped up startup on large repositories. Teams can now tune MCP tool-output truncation per repository through a .grok/config.toml file. The steady stream of refinements continues the momentum from Grok Build's earlier expansion into cloud sandboxes.
Grok 4.5 Does the Heavy Lifting
The quiet headliner is Grok 4.5 becoming the default model. Trained across tens of thousands of NVIDIA GB300 GPUs and developed alongside coding tool Cursor, Grok 4.5 is tuned for real software-engineering work and posts competitive scores on demanding benchmarks like Terminal Bench and SWE Bench Pro. Just as important for developers is efficiency: the model runs at roughly 80 tokens per second and solves tasks with about 4.2 times fewer output tokens than comparable leading models.





