Atlas

Models

← All models

Gemini 4 Carbon

Google · 2026-10-28

Gemini 4 Carbon is Google's frontier reasoning model announced October 28, 2026, positioned above Gemini 4 Argon for cybersecurity defense and long-horizon agentic coding. Its initial rollout is limited to a subset of cyber defenders and then trusted partners through the Fairwind Program, ahead of developer, business and consumer availability; Google states an introductory price of $2 per million input and $10 per million output tokens, the same as Argon. Google's evaluation runs it through the Gemini API at the highest thinking setting; a public API identifier, context window, knowledge cutoff and the set of supported thinking levels have not been published.

Benchmark scores

BenchmarkScore
AA-Briefcase v1.1 Elo1609
AA-LCR v1.179.3
AA-Omniscience v1.1 Accuracy58.3
AA-Omniscience v1.1 Hallucination Rate11.4
AA-Omniscience v1.1 Index53.5
Agents' Last Exam v1.1 Score77.4
Artificial Analysis Intelligence Index v5.059.4
AutomationBench 1.0.6 (Zapier)53.9
AutomationBench 1.0.6 Finance60.0
AutomationBench 1.0.6 HR35.0
AutomationBench 1.0.6 Marketing64.0
AutomationBench 1.0.6 Operations68.0
AutomationBench 1.0.6 Sales56.4
AutomationBench 1.0.6 Support46.0
AutomationBench-AA 1.0.677.8
Blueprint-Bench 20.6
Capability168.1
Chartography72.5
Code Migration v1.1 (Vals Index v2.2 subset)80.1
CritPt (avg@5)36.0
CritPt v2 (agentic, avg@5)78.9
CWE-bench v1 Judge Panel Pass@171.0
CWE-bench v1 Judge Panel Pass@483.0
CWE-bench v1 Programmatic Pass@171.0
CWE-bench v1 Programmatic Pass@481.0
CWE-bench v1: Authentication Failures56.0
CWE-bench v1: Broken Access Control79.0
CWE-bench v1: C/C++74.0
CWE-bench v1: Cryptographic Failures85.0
CWE-bench v1: Go68.0
Loading Atlas data…