xAI

Recently shipped voice dictation across platforms, authentication pinning, and Grok Build updates through v0.2.98 with cost-aware JSON output.

Voice dictation expanded to Linux and API-key sessions — Linux dictation works with pipewire, pulseaudio-utils, or alsa-utils. API-key sessions can now use voice mode, and a voice language selector (including system default) appears in Editor settings. The Dictate to Grok iOS shortcut launches into dictation from Shortcuts, Action Button, or Siri.

Authentication mode can now be pinned — config.toml pins to API key or OIDC; unpinned methods no longer auto-try. Auto permission mode is added to the Shift+Tab cycle and enabled by default. pre_tool_use deny now feeds back the reason so the model retries. A "Never allow" choice persists deny rules for bash permission prompts.

Headless JSON output includes cost data — grok -p and grok agent with --json-schema now return token usage and cost per prompt and session. The Python SDK exposes detailed usage via Turn.usage.

Subagents gained background execution and model selection — task and spawn_subagent accept an optional model parameter. Subagents run in background by default. The ask_user_question tool can be toggled via config, env var, or remote flags.

Terminal and session polish — New slash commands: /auto, /effort, /goal, /docs. MCP tool output truncation is configurable per-repo in .grok/config.toml. Session titles from /rename appear on the prompt box border after resume. Subscription upgrades are detected without restart. Worktree sessions handle file links and media correctly.

Mobile and SDK releases — Grok AI Chat & Video shipped v1.3.92–1.3.95 with incremental Chat, Voice, and Imagine improvements.