Google Unveils Gemini 4 Carbon With Gains in Long-Running Coding and Cyber Defense

Carbon is Google's new flagship model, and the company says it handles long engineering jobs that can run for days, such as tracing and patching bugs across large codebases. In internal tests it reduced successful cheating on Terminal-Bench 4.0 to 0.0 percent, cut unsupported claims by about half compared with its predecessor Argon, and resisted every attack in an indirect prompt injection benchmark. Google also reports a 26 percent drop in overreach and a 29 percent drop in under-service on multi-step plans. The first users are a cohort of cyber defenders, and the company says developers, businesses and consumers will get access later, without giving a date. For founders building agents, the useful signals are the reliability and injection resistance numbers. Teams should watch how Carbon's pricing and access terms arrive before planning any product changes around it.