Overview
- The company launched Grok 4.7 on Monday, Sept. 21, and made it available across the Grok app, Cursor, Grok Build and the xAI API at the same starting token price as Grok 4.6 ($2 per million input, $6 per million output).
- Grok 4.7 uses a larger base model reported at about 2.1 trillion parameters and was trained with a longer reinforcement‑learning run weighted toward harder, multi‑hour tasks to improve long‑context handling and self‑verification.
- xAI published benchmark gains — including higher scores on CursorBench, Terminal‑Bench, DeepSWE and AA Briefcase — that narrow gaps with frontier systems but independent leaderboards show Grok still trails top models on endurance and reliability for many multi‑hour agent runs.
- The release includes a rebuilt safeguard stack that xAI says tops LatchBio and HackerBench metrics for biosafety and jailbreak resistance, and the company is granting invite‑only red‑team access to select cybersecurity partners for defensive testing.
- Early product signals include Grok Bot reaching roughly 418,000 weekly users and broad availability in third‑party tooling, and observers warn that harness‑specific training may boost agent performance while making models harder to swap between different tool chains.