Google announces Gemini 4 Carbon with leading cybersecurity and coding benchmarks
Ars Technica|AIDEN Editorial|Oct 28, 2026
Google has unveiled Gemini 4 Carbon just two weeks after releasing Argon, claiming the top position on independent intelligence benchmarks. General access remains restricted to vetted defense specialists due to the frontier model's advanced exploit creation capabilities.
Google has officially announced its next frontier AI model, Gemini 4 Carbon, arriving two weeks after the release of Gemini 4 Argon. Koray Kavukcuoglu, senior vice president and chief AI architect at Google DeepMind, stated via the Google Blog that the model builds on the technical base of Argon through researcher collaboration, driving performance gains alongside enhanced safety guardrails. The rollout represents an aggressive move to regain ground in coding and autonomous agents from rival frontier systems including Anthropic Claude Opus 5.5 and OpenAI GPT-6 Astra.
Evaluation data highlights significant benchmark progress across intelligence and development metrics. On the Artificial Analysis Intelligence Index v5.0, Gemini 4 Carbon recorded an overall score of 59, climbing nine points above Argon to edge past Claude Opus 5.5 at 58 and GPT-6 Astra at 57. The model also reached 1,769 points on WebDev Arena, an 89-point gain that lifted it to fourth place on the leaderboard. In specialized cybersecurity testing, Carbon achieved an 88.2 percent score on SEC-bench Pro, 54.8 percent on ExploitGym, and 71.0 percent on the CWE-bench patch automation test, while proving twice as likely as Argon to write vulnerability-free code. Meanwhile, Google internal Chrome evaluations indicated that Carbon generated high-quality patches at a rate 10 times better than competing non-Google models, though it trailed Claude Code by three points on the Artificial Analysis Coding Agent Index.
Because its advanced vulnerability detection and automated exploit mechanisms present dual-use risks, Google has withheld a broad general release. The company is engaging in pre-release voluntary testing with the United States federal government and limiting initial model access to authorized defensive cybersecurity specialists via the Fairwind Program. Wider API endpoints for regular developers and enterprise clients remain postponed pending further guardrail refinements. Google is offering promotional pricing matching Argon at $2 per million input tokens and $10 per million output tokens, alongside a 95 percent discount on cached inputs, but standard rates will eventually double to $4 and $20 respectively once the promotional window ends.
— AIDEN Editorial Team
What this means for the market
Google rapid introduction of Gemini 4 Carbon underscores an accelerating competitive cycle among frontier model developers vying for dominance in agentic workflows and automated software engineering. However, the decision to withhold immediate commercial endpoints highlights how dual-use cyber capabilities are increasingly constraining deployment schedules across the enterprise sector. As regulatory scrutiny over offensive software generation intensifies globally, cloud providers must balance headline benchmark gains against the operational realities of restricted rollouts and looming token price increases.
How this issue is unfolding
Google decision to unveil its superior frontier model Carbon just two weeks after rolling out Gemini 4 Argon reflects a calculated strategy to swiftly reclaim leadership in coding and autonomous agents, where it had fallen behind Anthropic Claude Opus 5.5 and OpenAI GPT-6 Astra. However, because its potent vulnerability discovery and automated exploit capabilities present a double-edged sword, Google opted for limited distribution through the Fairwind Program alongside voluntary pre-release testing with the United States federal government. Consequently, a recurring pattern has emerged in which benchmark victories are heavily promoted before general developers or enterprise customers can obtain access, prompting ongoing scrutiny over delayed external verification. At the same time, Google is defending its enterprise positioning by temporarily freezing token rates to counter persistent doubts regarding the real-world efficiency of its coding tools.
