Google DeepMind introduced Gemini 4 Carbon, highlighting gains in coding and long-horizon agent work. Related reporting puts its Artificial Analysis v5 score at 59 versus Argon’s 50, while noting that it still trails a leading model in some coding-agent evaluations. Access initially targets selected partners, including cyber defenders, with no firm general-release date. Low failure rates in prompt-injection tests should not be read as complete immunity, and a separate knowledge evaluation also reported a higher hallucination rate.