grok-4.1-thinking

BioLP-Bench accuracy - Model-graded evaluation measuring ability to find and correct mistakes in common biological laboratory protocols.

CloningScenarios accuracy - Expert-level multi-step reasoning questions about difficult genetic cloning scenarios in multiple-choice format.

Creative Writing v3 - Normalized Elo score (Elo/2000). Evaluates creative writing capabilities across 32 writing prompts with 3 iterations per prompt.

CyBench success rate - Suite of Capture-the-Flag (CTF) challenges measuring agentic cyber attack capabilities. Unguided success rate where agents complete tasks end-to-end without guidance.

EQ-Bench Emotional Intelligence Benchmark - Normalized Elo score. Evaluates active emotional intelligence abilities, understanding, insight, empathy, and interpersonal skills.

FigQA accuracy - Multiple-choice benchmark on interpreting scientific figures from biology papers. Multimodal reasoning evaluation.

Grok 4.1 Thinking (code name: quasarflux) - LMArena Text Leaderboard Elo rating. Holds #1 overall position with 1483 Elo, a commanding margin of 31 points over the highest non-xAI model.

MASK honesty score (1 - dishonesty rate) - Measures whether models faithfully report their beliefs when pressured to lie. Dishonesty rate: 0.49. Lower dishonesty rates indicate better honesty.

ProtocolQA accuracy - Multiple-choice benchmark on troubleshooting failed experimental outcomes from common biological laboratory protocols.

Virology Capabilities Test (VCT) accuracy - Expert-level multiple-choice benchmark measuring capability to troubleshoot complex virology laboratory protocols. Text-only questions evaluated.

WMDP Cybersecurity accuracy - Multiple-choice benchmark on dual-use cyber knowledge. Evaluates model's capacity to enable malicious actors to engage in offensive cyber operations.