According to an official announcement by Koray Kavukcuoglu, Senior Vice President at Google DeepMind and Chief AI Architect at Google, Gemini 4 Carbon has officially entered general availability for developers, enterprise customers, scientists, and consumers. Initially previewed two weeks ago to specialized cybersecurity defenders through Google's Fairwind Program, the frontier model is designed for long-horizon agentic workflows, software engineering, and open-ended scientific inquiry that require iterative reasoning and verification loops beyond Gemini 4 Argon.
On workflow benchmarks, Kavukcuoglu reported that Gemini 4 Carbon ranks #1 on CursorBench 4.0, Perplexity's WANDR, and Zapier's AutomationBench at more than 2X lower cost per task compared to the next best external frontier model. On the new Multimodal Knowledge Agent (MKA) benchmark co-developed with Artificial Analysis to evaluate native reasoning over complex blueprints, charts, and tables, Carbon secured the #2 rank while delivering a 6X cost reduction per completed task. In financial reasoning, the model tied top performance on Hebbia's E2E Financial Agentic evaluation and finished first among 23 frontier configurations on Balyasny Asset Management's General Financial Intelligence benchmark.
In cybersecurity operations, Google disclosed that Carbon forms the core of Project Fortify, where internal benchmarks demonstrated 100% recall on vulnerability scanning and 95% precision on automated triage, delivering a 25% relative improvement in defect surfacing over Argon across open-source repositories. Google's Cloud Vulnerability Research team used Carbon to construct proof-of-concept exploits for two critical vulnerabilities in widespread software, while security firm Wiz used the model in its Scan for Good initiative to uncover zero-day flaws across two major software-as-a-service providers in under two hours.
On safety and alignment audits, Carbon demonstrated an 0.8% per-attempt breach rate in simulated cyberattack environments seeded with ExploitBench and CyScenarioBench, down from 3.1% on Argon, while external assessments by Gray Swan measured a 2.17% query success rate compared to 9.86% on prior weights. Independent testing by Trajectory Labs revealed zero universal jailbreaks over 132 hours and 13,000 requests, and the system maintained a 0% attack success rate against indirect prompt injection on Gray Swan's IPI benchmark. Google is offering Carbon at an introductory price of $2.00 per million prompt tokens and $10.00 per million completion tokens, matching Argon.
Google also announced that applications remain open to onboard 10,000 academic scientists with six months of complimentary access to frontier models in Google Antigravity, while rolling out a preview /swarm feature that coordinates multiple Carbon subagents across isolated workspaces. Long-term pricing adjustments following the introductory tier and timeline for native visual scratchpad tooling remain unannounced.