Google DeepMind Launches Gemini 4 Carbon, Tops Artificial Analysis Intelligence Index v5

Google DeepMind released Gemini 4 Carbon just two weeks after its predecessor Argon. The new model scores 59 points on the updated Artificial Analysis Intelligence Index v5, surpassing Argon’s 50 and taking the overall top spot. Gemini 4 Carbon shows broad gains across agentic tasks, programming, scientific reasoning, multimodal understanding, and full‑stack web development. Google highlights its ability to handle long‑horizon engineering problems, including multi‑day coding sessions, root‑cause analysis, and subtle bug fixes in large codebases. In internal benchmarks, Carbon leads 15 of 16 tests, excels in long‑context and security tasks, and reduces prompt‑injection success from 0.7% to 0.0% on Gray Swan’s benchmark. Independent evaluations confirm its high reasoning tier performance, with top scores in complex reasoning, knowledge retrieval, and scientific tasks. While it still trails Claude Opus 5.5 in agentic coding by three points, it achieves significant gains in presentation and analytical quality on AA‑Briefcase, and its web‑development Elo rises by 89 points to 1,769, placing it fourth in the WebDev Arena.