J.Muck

Allgemein

Dev-Logs

📁 Projekte

✅ Aktiv

⏸ Stabil

📥 Neu

🧠 xAI Wiki

Entities

Concepts

Comparisons

🌐 Jamgrid Wiki

Impressum
Vault › Wiki › concepts › grok-models

Grok Models

Overview

Grok is the family of large language models developed by xai-org. Named after Robert Heinlein's "Stranger in a Strange Land" — "to grok" = to understand deeply. The current flagship is Grok 4.3 (released April 30, 2026).

Defining edge: Only major model with native real-time X (Twitter) data stream. Defining limitation: Calibration — high answer rate but high error rate when uncertain.

Active Model Variants (2026)

VariantReleasedContextMultimodalAPI In/Out (per M tokens)
---------
----------------------------------------------------------
Grok 4.3 (flagship)
2026-04-301Mtext, image, video$1.25 / $2.50
Grok 4.20 Reasoning
2026-03-312Mtext, image$2.00 / $6.00
Grok 4.20 Non-Reasoning
2026-03-312Mtext, image$2.00 / $6.00
Grok 4.1 Fast
2025-11-192M—$0.20 / $0.50
Grok 4 / Heavy
2025-07-09256K—$3.00 / $15.00
Grok 4 Fast
2025-09-192M—$0.20 / $0.50
Grok 3 / Mini
2025-02-17131K—

Architecture Notes

  • Grok-2: 55B params, 64 layers, 8 experts (MoE), 2 experts per token, 128K context, RoPE, GELU activation[grok-2-architecture]
  • Grok 3: Introduced DeepSearch, Think, RL at post-training scale
  • Grok 4: RL at pretraining scale (10× compute), multi-agent Heavy, native voice/camera, 256K context
  • Grok 4.20: Multi-agent "Society of Mind" (4 agents), 17% hallucination on AA-Omni (best in family)
  • Grok 5 (planned Q2 2026): 6T-10T parameters (per Musk); timing has slipped from Q1

Key Features

  • Real-time X data access (tweets, threads, trends)
  • Always-on reasoning (since Grok 4)
  • Image generation (Aurora)
  • DeepSearch (research mode)
  • Voice mode (native camera on mobile)
  • Function calling, structured output
  • Up to 2M token context (Fast variants, API)

Benchmarks

  • AA-Omni hallucination: Grok 4.20 = 17%, Grok 4 = 64%, Grok 4.1 Fast = 72%
  • HLE: Grok 4 Heavy = 50.7%
  • AIME: Grok 4 Heavy = 100%
  • LMArena: Grok 4.1 Thinking = 1483 Elo (#1 Text Arena)
  • AA Intelligence Index: Grok 4 = 73 (o3 = 70, Gemini 2.5 Pro = 70, Claude 4 Opus = 64)

Upcoming

  • Grok 4.4 (~1T params): 2-3 weeks from late Apr 2026
  • Grok 4.5 (~1.5T params): 4-5 weeks from late Apr 2026
  • Grok 5: Q2 2026 (volatile timeline)

> See comparisons/grok-pricing-tiers for full pricing breakdown.