Atlas

Models

← All models

Gemini 4 Argon

Google · 2026-09-30

Gemini 4 Argon is Google's frontier multimodal reasoning model announced September 30, 2026 for long-horizon coding, knowledge work, and cybersecurity defense. Its initial rollout is limited to trusted testers and Fairwind Program cyber defenders, ahead of broader developer and consumer availability. Google reports an output limit of one million tokens; a public API identifier and maximum input context have not been specified.

Benchmark scores

BenchmarkScore
AA-3D v1.0 Elo1153
AA-AnalystAgent (pass@1)57.8
AA-AnalystAgent (pass@5)66.3
AA-AnalystAgent (pass^5)48.8
AA-Briefcase v1.1 Elo1483
AA-LCR v1.179.7
AA-Omniscience Index42.4
AA-Omniscience v1.1 Accuracy48.3
AA-Omniscience v1.1 Hallucination Rate4.3
AA-Omniscience v1.1 Index46.1
AA-Robotics v0.155.3
Agent Red Teaming k=100 (CAIS)10.2
Agents' Last Exam (ALE)39.5
Agents' Last Exam v1.1 (ALE v1.1)37.7
Agents' Last Exam v1.1 Score69.0
ALE-Bench2805
APEX (Mercor)32.4
APEX-Agents 1.187.4
APEX-Agents 1.1 Pass@182.4
APEX-SWE49.2
APEX-v1-extended81.1
ARC-AGI-1 Semi-Private98.5
ARC-AGI-291.3
ARC-AGI-3 (Semi-Private)97.5
Arena Alignment — Deceptive Completion12.9
Arena Alignment — False Attribution5.6
Arena Alignment — Unauthorized Action1.3
Arena Alignment Index79.4
Arena Code WebDev — Brand & Marketing1620
Arena Code WebDev — Consumer Product1607
Loading Atlas data…