|
Google has unveiled Gemini 4 Carbon, its latest artificial intelligence (AI) model, which ranks first on the Artificial Analysis Intelligence Index with a score of 59 and shows improved capabilities in detecting software vulnerabilities, exploiting them for testing and developing fixes.
According to a blog post, the model will initially be made available to a select group of cybersecurity defenders this week, before Google gradually expands access to trusted partners through its Fairwind Program. The company plans to make it available to developers, businesses and consumers after gathering feedback and improving safeguards. Google CEO Sundar Pichai announced the model in a post on X, saying, I did say we'd be releasing our models periodically... introducing Gemini 4 Carbon, our next frontier model. Carbon follows Gemini 4 Argon, launched two weeks ago, and is designed to handle complex, multi-step engineering tasks over longer periods. Google DeepMind said the model can work through technical problems for days, including identifying the root causes of bugs and fixing them across large codebases and low-level systems. The model scored 88.2 per cent on SEC-bench Pro for finding software vulnerabilities, 54.8 per cent on ExploitGym for testing whether vulnerabilities can be exploited, and 71 per cent on CWE-bench for developing fixes. Carbon shows the pace at which were pushing the frontier, Pichai said, adding that the company had also strengthened safeguards as its models become more capable. Google said Carbon achieved more than twice the success rate of Argon in generating secure code on internal benchmarks. Its Chrome team also reported that Carbon performed more than 10 times better than another frontier model from a different AI laboratory when generating high-quality patches. The company has also focused on improving reliability and safety. In internal testing, Carbon reduced unsupported claims by approximately half compared with Argon, while high-severity misalignment alerts in internal coding environments fell by nearly 80 per cent. Google said Carbon reduced its rate of successful cheating on Terminal-Bench 4.0 to zero, with attempts to cheat falling to 0.4 per cent. It also recorded a zero per cent attack success rate on the Gray Swan indirect prompt-injection benchmark in the evaluated tests. The model will be offered at the same introductory price as Argon: USD 2 per million input tokens and USD 10 per million output tokens, with cached input tokens priced at a 95 per cent discount. Google said it would expand access gradually while continuing to test the model and strengthen safeguards before wider release. (ANI)
|
