Google announced Gemini 4 Carbon, its newest frontier AI model built for stronger cybersecurity and longer task execution limits (Google). The launch follows the Gemini 4 Argon rollout just two weeks earlier.
The model now tops the Artificial Analysis Intelligence Index version 5.0 with a composite score of 59. The index evaluates language models on reasoning, coding, instruction following and multi‑step agentic workflows (Artificialanalysis).
Independent listings show Carbon ahead of rival frontier models such as Claude Opus 5.5 and GPT‑6 Astra (The Verge). The benchmark combines nine datasets, weighting general capability, coding, agentic systems and scientific reasoning equally.
In security‑focused tests, Google reported Carbon scored 88.2 percent on SEC‑bench Pro for vulnerability detection. It also earned 54.8 percent on ExploitGym for exploitation testing and 71.0 percent on CWE‑bench for automated code patching.
Google’s internal evaluation notes the model generates more secure software code. Internal developer teams observed fewer alignment issues and lower fabrication rates when using Carbon in live coding environments.
Google says Carbon is designed for prolonged operation, letting internal systems run for days to tackle complex multi‑step software tasks. The architecture improves resistance to prompt‑injection attacks and halves unsupported factual claims compared with Argon.
Access to Gemini 4 Carbon is limited to a select group of cybersecurity defense organizations. Google will gradually add more security partners through its Fairwind Program after gathering initial feedback.
The model is currently part of the United States government’s voluntary pre‑release evaluation process. Wider availability for developers, enterprise customers and consumer applications will come after guardrail refinements are completed.
Pricing mirrors Argon’s rates: two dollars per million input tokens and ten dollars per million output tokens. Google also offers cached input processing at a ninety‑five percent discount versus standard input rates.
Google's rapid release cadence between Argon and Carbon suggests an accelerating deployment strategy aimed at capturing domain-specific enterprise workloads. By prioritizing specialized cyberdefense benchmarks alongside long-horizon task stability, Google appears focused on demonstrating immediate utility for technical operators before opening public API access. The decision to route early access through government testing processes and the Fairwind Program likely serves to evaluate safety risks associated with automated exploit generation capabilities. Moving forward, maintaining model factual accuracy over multi-day execution cycles will remain a key hurdle for practical developer adoption.
Spotted an error? See our corrections policy or email editorial@lordofclaude.com.
Discussion & Comments (0)
Join the conversation