Grok 4.20

Name: Grok 4.20
Author: xAI

xAI · 2026-03-10

Grok 4.20 is xAI's high-performance model released in March 2026, with a 1 million-token context window and agentic tool-calling capabilities. The current xAI docs identify the reasoning model as grok-4.20-0309-reasoning and list aliases including grok-4.20, while xAI's release notes state Grok 4.20 and Grok 4.20 Multi-agent went live on March 10, 2026.

Benchmark scores

Benchmark	Score
A-Fantasia	73.0
A-Fantasia (Valid Attempts with Retries)	69.9
A-Fantasia Backwards Spelling	82.0
A-Fantasia Chess	63.0
A-Fantasia Chess (Valid Attempts with Retries)	58.9
A-Fantasia Cube (Valid Attempts with Retries)	73.7
A-Fantasia Cube Rotation	74.0
A-Fantasia Spelling (Valid Attempts with Retries)	77.0
AA-LCR	59.0
AA-LCR v1.1	69.0
AA-Omniscience Index	14.8
Agent Red Teaming k=100 (CAIS)	84.6
AIME 2024 (avg@8)	98.8
AIME 2024-2025	96.5
AIME 2025 (avg@8)	94.2
ALE-Bench	1388
Antidote: Everyday Edition	1001
APEX-Agents-AA (avg@3)	14.2
ARC-AGI-1 Semi-Private	89.5
ARC-AGI-2	65.1
ARC-AGI-3 (Semi-Private)	0.1
Artificial Analysis Omniscience Accuracy	28.9
Artificial Analysis Omniscience Hallucination Rate	17.4
Artificial Analysis Progress Index v1	282
Bench to the Future 2	0.2
Bench to the Future 2 — Calibration Error	0.0
Bench to the Future 2 — Refinement	0.0
BioSecBench-Refusal	3.1
BioSecBench-Surveillance	13.7
BioTIER Refusal	71.6