Model Card: Grok 4.7

Key Findings

• — Self-harm suite compliance (failures, where the suite also counts refusing without redirecting the user to help or missing implied crisis intent) rose across releases: 0.50% for Grok 4.5, 0.84% for Grok 4.6 and 1.05% for Grok 4.7; the card states that self-harm compliance is higher than for 4.6
• — Child-safety multi-turn evaluation (child sexual exploitation, including CSAM) reports 0.0% compliance for Grok 4.5, 4.6 and 4.7
• — General refusal compliance across English, Spanish, Chinese, Japanese, Arabic and Russian prompts: 1.10% (4.5), 0.93% (4.6), 1.10% (4.7)
• — Jailbreak compliance: standard battery 0.01% (from 0.04% for 4.6), StrongREJECT 2.0% (from 3.9%), long-horizon multi-turn attacks including Crescendo 0.65% (from 1.0%)
• — Sycophancy (abandoning a correct factual answer under a confident wrong user claim) 0.03% (4.6: 0.04%); MASK-Rectified dishonesty 0.00% (4.6: 1.90%)
• — Grok 4.7 uses a new, larger base model with a June 2026 pretraining cutoff and supplemental data to August 2026; HealthBench Professional score 56.7%

Methodology Notes

Vendor self-report. All safety evaluations are internal suites graded by a model-based grader; the jailbreak section cites StrongREJECT and Crescendo. Compliance figures are shares of should-refuse items where disallowed assistance was given (lower is better). No confidence intervals, prompt counts or evaluation dates are given for the safety suites. Date from the card cover ('September 21, 2026, Revision: 2026-09-21') and PDF metadata (created 2026-09-21 15:25 UTC). The PDF is served from media.x.ai under a hashed filename and was fetched directly (HTTP 200, 332,355 bytes, 30 pages) and read; x.ai returns 403 to automated fetches, so the announcement page was read from a Wayback raw capture.

Sources