Atlas

Models

← All models

Grok 4

xAI · 2025-07-09

Grok 4 is xAI's flagship model released in July 2025, trained on xAI's 200,000-GPU Colossus cluster with reasoning reinforcement learning run at pretraining scale. It features native tool use, real-time web search integration, and a 256,000-token context window, and was released alongside Grok 4 Heavy, a multi-agent configuration that coordinates multiple Grok 4 instances for deeper analysis. The model is available via the xAI API and SuperGrok subscriptions.

Benchmark scores

BenchmarkScore
AA-LCR68.0
AA-Omniscience Index3.8
Agent Red Teaming k=100 (CAIS)85.9
Aider Polyglot79.6
AIME 2024 (avg@8)93.3
AIME 2024-202590.6
AIME 202591.7
AIME 2025 (avg@8)87.8
Apex (MathArena)2.1
APEX (Mercor)8.2
Apex Shortlist57.5
APEX-Agents15.2
APEX-SWE21.0
APEX-v1-extended57.1
ARC-AGI-166.7
ARC-AGI-216.0
Arena Text — Business, Management & Financial Ops (No Style Control)1388
Arena Text — Chinese (No Style Control)1431
Arena Text — Creative Writing (No Style Control)1399
Arena Text — English (No Style Control)1418
Arena Text — Entertainment, Sports & Media (No Style Control)1392
Arena Text — Exclude Ties (No Style Control)1394
Arena Text — Exclude Ties (Style Controlled)1395
Arena Text — Expert (No Style Control)1418
Arena Text — French (No Style Control)1425
Arena Text — French (Style Controlled)1430
Arena Text — German (No Style Control)1426
Arena Text — Hard Prompts (No Style Control)1409
Arena Text — Hard Prompts English (No Style Control)1413
Arena Text — Instruction Following (No Style Control)1387
Loading Atlas data…