Home · Technology · Nov 3 archive
Google's Gemini 4 Carbon Outperforms Competitors in Cybersecurity and Coding
· Confirmed
In Short: Google's Gemini 4 Carbon model excels in cybersecurity and coding benchmarks, outperforming competitors like Claude and GPT-6 Astra.
Google's Gemini 4 Carbon, announced on October 28, has been benchmarked by SemiAnalysis in their GDPx test suite, which includes 206 tasks across various domains. The model has shown significant improvements over its predecessors, though a public release date has not been confirmed.
In the GDPx benchmark, Gemini 4 Carbon outperformed GPT-6 Astra, which scored lower despite costing more at about $54 per task. Gemini 4 Carbon's cost was just over $16 per task, making it more cost-effective than both GPT-6 Astra and Claude, which cost around $51 per task.
According to SemiAnalysis, Gemini 4 Carbon's performance aligns well with their subjective impressions of the model. The benchmark leaderboard includes every major frontier model from companies like OpenAI, Anthropic, and Meta, detailing both scores and costs.
Gemini 4 Carbon's cybersecurity capabilities are particularly noteworthy. Internal tests showed Carbon had a 0.0% attack success rate, followed by Argon at 0.7%, and Claude Opus 5.5 at 1.0%. This performance was validated on external benchmarks such as SEC-bench Pro, ExploitGym, and CWE-bench.
Google claims that Carbon performs more than 10 times better than another lab's large frontier model when generating high-quality patches. It also leads the Artificial Analysis Intelligence Index with a score of 59, surpassing Anthropic's Claude Opus 5.5 (58) and OpenAI GPT-6 Astra (57).
In addition to cybersecurity, Carbon excels in coding, agentic tasks, and full-stack web development. It scored 77.9% on DeepSWE v1.1, a benchmark for software engineering, and led in knowledge work, long documents, and reading charts and video.
Google's goal with Gemini 4 Carbon is to deliver meaningful improvements in software engineering and cybersecurity defense without compromising on safety or reliability. The model will initially be available to a subset of cyber defenders at $2 per million input tokens and $10 per million output tokens, with rates rising to $4 and $20 after the introductory period.
Gemini 4 Carbon's enhanced capabilities are expected to benefit developers, businesses, and consumers, though access will be limited to select vetted cybersecurity teams initially.
Background
Google announced Gemini 4 Carbon, a new AI model designed to excel in cybersecurity and complex coding tasks.
Google positions Carbon as the top of the Gemini 4 line, emphasizing its capabilities in coding, agentic tasks, and full-stack web development.
What's confirmed
- "significant leap in cybersecurity capabilities."
What's still developing
- The latest AI model to come from the minds over there at the Mountain View-based firm is called Gemini 4 Carbon.
- Carbon is said to outperform most other models, capturing the number one spot on version 5.0 of the Artificial Analysis Intelligence Index.
- However, cybersecurity is where Carbon really shines.
- While Carbon builds on the same safety and security stack as Argon, Google went a step further to also improve the model’s alignment.
- Recommended by Our Editors Claude: Stop Using Me to Recreate Games Facial Recognition Could Be Coming to a Starbucks Near You Sam Altman Suggests Running AI Locally Is a Waste of Memory Carbon will initially be available to “a subset of cyber defenders” at $2 per million input tokens and $10 per million output tokens.
- "GPT‑6.1 Terra meets the Critical threshold in cybersecurity under our Preparedness Framework, and maintains the protections developed for GPT‑6 Astra," OpenAI said.
- Google launched Gemini 3 on November 18, 2025 and said Gemini 3 Pro led LMArena at the time.