Google Announces Gemini 4 Carbon with Focus on Cybersecurity and Long-Horizon Tasks
· GlobeWire Daily Editorial Desk · Source: blog.google

Image: Google · Photo source · Resized and letterboxed to 16:9; no AI generation
Google has announced Gemini 4 Carbon, described as its strongest model to date, featuring advanced cybersecurity capabilities and support for long-horizon engineering tasks. According to the company, the model is designed to tackle complex, multi-step workflows.
Performance and Benchmarks
Google reports that Gemini 4 Carbon ranks high on several industry benchmarks:
- SEC-bench Pro: 88.2% for vulnerability discovery.
- ExploitGym: 54.8% for exploitation tasks.
- CWE-bench: 71.0% for code patching.
- Artificial Analysis Intelligence Index (v5.0): Ranked #1 with a score of 59.
Rollout and Pricing
Google is adopting a deliberate rollout strategy, engaging with the U.S. government's voluntary pre-release testing process. Initial access begins this week for a subset of cyber defenders and will expand gradually through the Fairwind Program. Availability for developers, businesses, and consumers via product surfaces is planned for a later date following feedback integration.
The model shares the introductory pricing structure of Gemini 4 Argon:
- Input tokens: $2 per million ($4 after the introductory period).
- Output tokens: $10 per million ($20 after the introductory period).
- Cached input tokens: 95% discount off the input price.
Safety and Alignment
Carbon builds upon the multi-layered safety stack of its predecessor, Argon. Google highlights improvements in reducing misalignment, lowering fabrication rates, and achieving a 0.0% success rate against evaluated attacks on the Gray Swan Indirect Prompt Injection Benchmark.
Verification
Readers and developers can verify these announcements by checking official Google DeepMind communications and tracking updates regarding the Fairwind Program and public product releases.
AI-assisted reporting; source claims are attributed and have not been independently verified.



