Grok

xAI Grok

xAI

xAI's Grok 4.20 boasts a 2M token context window — the largest among frontier models — with multi-agent architecture, real-time X data integration, and the lowest hallucination rate at 78% factual accuracy.

grok.com
Last updated: April 20, 2026
$2/M
4.20 Input
2M
Context Window
30K
Max Output
182 t/s
Speed
★★★★
Coding
★★★★★
Reasoning
★★★★
Writing
★★★★★
Speed

Available Models

ModelInput $/1MOutput $/1MContextBest For
Grok 4.20 (flagship)$2.00$6.002MComplex reasoning, research, agents
Grok 4.1 Fast$0.20$0.502MSpeed-first reasoning, search
Grok 4$2.00$6.002MPrevious gen, still available
Grok 3 Mini$0.30$0.50131KLightweight tasks

Consumer Plans

PlanPriceIncludes
Free$0/moLimited Grok access via X
SuperGrok$30/moEnhanced access, DeepSearch
X Premium+$40/moFull X features + Grok
SuperGrok Heavy$300/moHighest limits, priority

Strengths & Weaknesses

Strengths
  • Lowest hallucination rate: 78% factual accuracy (Artificial Analysis)
  • Multi-agent architecture (Grok, Harper, Benjamin, Lucas)
  • Real-time search and X data integration
  • Massive 2M context window at competitive pricing
  • 182 tokens/sec — among fastest reasoning models
Weaknesses
  • Trails Claude on complex coding (SWE-bench) (Awesome Agents)
  • Potential bias toward xAI/Musk-adjacent topics
  • ~15s reasoning time to first token on complex queries

Best For

  • Real-time information synthesis and fact-checking
  • Financial analysis and research
  • Agentic tasks requiring low hallucination
  • Long-context processing (2M window)
  • Speed-critical production workloads

Latest Release NEW

Grok 4.20 Beta 2 — April 2026

  • Multi-agent system with enhanced step-back reasoning
  • Industry-leading speed and low hallucination rates
  • Trained on the Colossus supercluster — largest AI training cluster
  • Native tool use and vision capabilities
  • 2M token context window — largest commercially available

Previous Releases

Grok 4.20 Beta — February 2026

  • Initial beta with multi-agent reasoning architecture
  • Introduced step-back reasoning for complex problem solving
  • Integrated with X platform for real-time information access

Grok 3 — December 2024

  • First Grok model trained on Colossus supercluster
  • Major improvements in math, coding, and reasoning
  • Set the foundation for the 4.x multi-agent architecture