Grok 4.5 Independent Benchmarks Reveal 54% Hallucination Rate — Double Its Predecessor

What happened

The day after Grok 4.5 launched (July 8, 2026), independent firm Artificial Analysis published benchmark results, revealing a significant rise in hallucination rates despite accuracy gains.

Context and impact

Grok 4.5 ranked 4th on the Artificial Analysis Intelligence Index (score 54) behind Fable 5, GPT-5.5, and Opus 4.8. The model leads all tested models on agentic tool-use, making it strong for sequential tool-call workflows and automation. It's also the cheapest per Intelligence Index task at $0.31. The high hallucination rate — the model is 'confidently wrong' — is a warning for fact-critical deployments.

Details

• Hallucination rate: 54% (Grok 4.5) vs. 25% (Grok 4.3) — a 116% increase
• Accuracy (AA-Omniscience): 52% (Grok 4.5) vs. 35% (Grok 4.3)
• Intelligence Index: score 54, 4th place overall
• Coding Agent Index: 76, significantly lower token usage than competitors
• Best agentic tool-use performance of any model tested
• Cost: $0.31 per Intelligence Index task — cheapest model tested