Overview
- Grok 4.7 went live on Monday and is available across the Grok app, Cursor, Grok Build, the xAI API and third‑party platforms at the same baseline price of $2 per million input tokens and $6 per million output tokens.
- xAI says the model uses a larger base model and a longer reinforcement‑learning run focused on harder, multi‑hour tasks to check its own work more often and manage longer context.
- The company published benchmark gains over Grok 4.6, including CursorBench 4.0 at 46.3% versus 40.4% and Terminal‑Bench 4.0 at 38% versus 20.3%, while noting mixed placement against rival frontier models on some tests.
- xAI rolled out an entirely new safeguard stack and published refusal and jailbreak metrics, reporting that HackerBench v0.3 allowed 3.3% of risky prompts and that Grok 4.7 scored 62.4% on LatchBio’s biosafety test, and it is giving select cybersecurity partners invite‑only red‑team access for defensive research.
- Independent reporting says Grok 4.7 runs on about 2.1 trillion parameters and used supplemental SpaceX engineering data in training but those scale and data claims come from press reports and have not been independently verified; the release continues xAI’s strategy of trading lower per‑token cost for broadly useful performance.