ai · · 2 min read

Google Unveils Gemini 4 Carbon With Major Coding And Security Gains

By Rachel Lin

Google Unveils Gemini 4 Carbon With Major Coding And Security Gains

Accelerated Development Cycle

Google officially launched Gemini 4 Carbon, its latest frontier model. This rapid update follows the Argon release by just a few weeks. The speed is attributed to a joint effort between researchers and models. They work together to accelerate technical capability development.

Technical Performance Improvements

Carbon records significant progress in key areas. In coding, web development, and agentic tasks, the model shows notable growth. Rather than guessing or forcing a path forward, Carbon retrieves relevant sources. It presents discrepancies to the user before acting. This creates a much better collaborative experience.

Reduced Hallucinations

In internal tests, Carbon cuts unsupported claims by roughly half compared to Argon. Instead of fabricating evidence or hesitating when information is missing, the model includes methodological warnings. It presents clearly what remains to be done. This ensures a more reliable output.

Exceptional Cybersecurity Results

Results in cybersecurity are outstanding. Carbon holds a firm advantage in SEC-bench Pro tests. It achieves an 88.2% rate for vulnerability identification. It scores 54.8% in ExploitGym for exploitation. It reaches 71.0% in CWE-bench for applying patches. On internal benchmarks, the model improves secure code generation. It reduces misalignment and fabrication rates in live coding environments.

Transparency And Honesty

Regarding transparency and honesty, Carbon was trained to respond correctly to complex tasks. Under pressure, it performs reliably. In internal tests, Carbon reduced overreach by 26%. It cut under-service by 29%. This ensures much more reliable execution of multi-step plans. The model respects project constraints and stays strictly within scope.

Availability And Pricing

Gemini 4 Carbon is described as soon available for developers, businesses, and consumers. The model will initially be made available to cybersecurity defendersthis week. For introductory pricing, Google offers a rate of $2 per million input tokens. Output tokens cost $10 per million. Cached input tokens are priced at a 95% discount. After the promotional period, costs will rise to $4 per million input tokens and $20 per million output tokens.

More stories:

Share:

Leave a comment

Comments are moderated. Yours will appear once approved. Maximum 2 comments per hour.