No retraining required; write your policy in English and the model evaluates text or images against it, matching classifiers up to 7x its size, Apache 2.0 on Hugging Face as of August 4.
Why it matters: A production-grade open-weight guardrail that teams can customize without ML expertise lowers the safety-layer barrier for content-facing agent deployments.
A new toggle (Ctrl+Alt+F) collapses tool activity into compact per-turn summaries with a live running-tool indicator, reducing context noise during long agentic sessions.
Why it matters: Developers building long agentic workflows get a cleaner signal on what the agent is doing without wading through raw tool logs.
grok-voice-latest flips to Think Fast 2.0 (0.70s TTFA, 2x transcription accuracy in noise, $0.08/min); callers wanting to stay on v1 must pin grok-voice-think-fast-1.0 before end of day.
Why it matters: A capable open alternative to OpenAI's Realtime API for voice-first agent builds — and an action item for any API caller using the latest alias.