Overview
- The model, released Monday, is live across Cursor, Grok Build, the Grok API, third‑party routers, the Grok app, X, and Tesla dashboards at the same starter price as Grok 4.6: $2 per million input tokens and $6 per million output tokens.
- xAI reports Grok 4.7 uses a larger base model (reported around 2.1 trillion parameters) and an extended reinforcement learning schedule focused on harder, multi‑hour tasks to improve self‑verification and long‑context handling.
- On vendor benchmarks published by xAI, Grok 4.7 improved over 4.6 on CursorBench (46.3% vs 40.4%) and Terminal‑Bench (38.0% vs 20.3%), while independent leaderboards place it competitive with but not consistently ahead of Anthropic and OpenAI on several tests.
- xAI rolled out a redesigned safeguard stack and reported top‑line biosafety and cyber‑refusal numbers, and it is offering invite‑only red‑team access to select cybersecurity partners for defensive research.
- Reviewers warn that, despite gains, Grok 4.7 still fails a majority of long‑running agent tasks in independent endurance tests, so users should expect better cost‑per‑task but continued reliability limits for multi‑hour autonomous workflows and agentic coding.