Google has launched its new flagship model, Gemini 4 Carbon, which tops the Artificial Analysis Intelligence Index 5.0 with 59 points, surpassing Claude Opus 5.5 and GPT-6 Astra. The model closes the capability gaps left by its predecessor, Gemini 4 Argon, and can operate autonomously for multiple consecutive days to complete complex engineering tasks. It delivers outstanding performance on rigorous tests such as terminal engineering, long-context multi-step decision-making, and cybersecurity offense and defense, falling only slightly behind competitors on a few pure code engineering metrics. Its token consumption efficiency is higher, with debut pricing at just half that of Claude Opus 5.5, cached input discounts as low as $0.2 per million tokens, and actual usage costs under 40% of competitors'. Meanwhile, the model significantly reduces cheating and hallucination rates, with out-of-bounds and perfunctory operations markedly decreased, achieving 100% immunity to prompt injection. Google leverages its RSI system to advance multiple model checkpoint iterations in parallel, releasing an upgraded version within two weeks. Currently, owing to its strong vulnerability discovery and exploitation capabilities, Carbon will first be made available to select cybersecurity defense personnel.