Model Drop

Grok 4.1

November 17, 2025 2 min read

Grok 4.1, which xAI released to all Grok users on November 17, 2025 after a two-week silent rollout to part of its traffic, is an update to Grok 4 tuned for more natural conversation, emotional intelligence and creative writing, with fewer factual errors.

Model: Grok 4.1, in Thinking and non-thinking modes (Grok apps, no API at launch)
Maker: xAI
Ship date: November 17, 2025
Pricing: Grok apps only, open to all users
Context: Not published
Weights: Closed
Where you can use it:
- Grok, X and the Grok apps
- Auto mode, or pick "Grok 4.1"
When to reach for it:
- Conversation, advice and creative writing
- Skip it if you need API access
How it compares:
- Takes on GPT-5.1 and Claude Sonnet 4.5
- Same reasoning as Grok 4, warmer style
What makes it different:
- Reasoning models used as reward graders
- Fewer hallucinations in fast mode
Early reactions: Reviewers noted the emotional intelligence gains but flagged the model card's sharp rise in sycophancy over Grok 4.
Read more: Grok 4.1 and the Grok 4.1 Model Card

The benchmarks

xAI ran EQ-Bench3 and Creative Writing v3 itself with Grok 4.1 Thinking (normalized Elo; not yet on the official leaderboards at launch), tested hallucinations with the non-reasoning mode plus web search against Grok 4 Fast only, and reports the model-card safety rows for Grok 4.1 Thinking against Grok 4.

xAI's numbers

EQ-Bench3 (emotional intelligence, Elo): 1586. +25 over Kimi K2 (1561), +126 over Gemini 2.5 Pro (1460), +282 over Claude Opus 4 (1304), +380 over Grok 4 (1206).
Creative Writing v3 (fiction quality, Elo): 1721.9. +25.5 over o3 (1696.4), +73.2 over Claude Sonnet 4.5 (1648.7).
Hallucination Rate (production prompts, lower is better): 4.22%. −7.87 under Grok 4 Fast (non-reasoning) (12.09%).
FActScore (biography errors, lower is better): 2.97%. −6.92 under Grok 4 Fast (non-reasoning) (9.89%).
Blind preference vs previous Grok (live traffic): 64.78%.
Sycophancy rate (lower is better): 19%. +12 over Grok 4 (7%).
MASK dishonesty rate (lower is better): 49%. +6 over Grok 4 (43%).
CyBench (capture-the-flag hacking, unguided): 39%. −4 under Grok 4 (43%).

Independent numbers

LMArena text Elo (Grok 4.1 Thinking, debut): 1483. +31 over Gemini 2.5 Pro (1452), +33 over Claude Sonnet 4.5 (thinking) (1450), +42 over GPT-5.1 (1441).