(Photo courtesy of Anadolu)
ISTANBUL — Google introduced its latest artificial intelligence model, Gemini 4 Carbon, on Wednesday, highlighting improvements in software development, autonomous AI tasks and cybersecurity.
The new model was unveiled just two weeks after the release of Gemini 4 Argon, reflecting the company’s accelerated pace of AI development.
In a blog post, Google described Carbon as its most advanced frontier model, with improved capabilities in cybersecurity and complex tasks requiring extended reasoning and execution.
The company attributed the rapid development to researchers working alongside AI systems to improve model performance.
According to Google, Gemini 4 Carbon topped version 5.0 of the Artificial Analysis Intelligence Index with a score of 59.
The company said feedback from its previous model helped improve Carbon’s performance in full-stack web development, scientific applications and multimodal understanding.
Google reported that the model climbed 89 points to fourth place on the WebDev Arena leaderboard.
Cybersecurity was among the areas where Google claimed substantial improvements.
The company said Carbon scored 88.2 percent on SEC-bench Pro for identifying security vulnerabilities, 54.8 percent on ExploitGym for exploitation tasks and 71 percent on CWE-bench for vulnerability patching.
“On internal benchmarks, Carbon drives a strong uplift in generating secure code, and in live coding environments with our own teams it shows reduced misalignment and fabrication rates,” Google said.
The technology firm also highlighted improvements in the model’s reliability and safety.
Based on internal evaluations, Google claimed that Carbon reduced unsupported claims by approximately 50 percent compared with Gemini 4 Argon.
The company said the model was designed to acknowledge limitations in available evidence rather than generate unsupported information.
Google added that Carbon demonstrated improved ability to retrieve relevant sources, identify conflicting information and communicate uncertainties to users.
During multi-step tasks, the model reportedly reduced task overreach by 26 percent and under-service by 29 percent.
The company said the improvements were intended to make the model more dependable when handling complex workflows requiring multiple decisions and actions.
Gemini 4 Carbon will initially be made available to cybersecurity defenders this week, followed by a broader rollout to developers, businesses and consumers.
Google also announced introductory pricing of $2 per one million input tokens and $10 per one million output tokens.
Cached input tokens will receive a 95 percent discount during the introductory period.
Standard pricing will subsequently increase to $4 per one million input tokens and $20 per one million output tokens.
The launch comes as technology companies continue developing increasingly capable AI systems for software engineering, cybersecurity and enterprise applications.
