Google's Gemini Carbon Set to Exceed 55% on HLE-Diamond Benchmark

Google’s highly anticipated Gemini 4 model, internally codenamed “Carbon,” is expected to make a formidable debut on the Humanity’s Last Exam (HLE-Diamond) benchmark, with early indicators pointing to an accuracy score of 55% or higher. The model, which has been rigorously tested by Google employees separate from the previously released Gemini 4 Argon, represents a significant leap in advanced AI reasoning. If the score holds upon its official listing, Carbon will establish a new standard for high-reasoning capabilities on the rigorous 1,000-question HLE-Diamond dataset.

The HLE-Diamond benchmark, maintained by the Center for AI Safety, has emerged as a critical proving ground for evaluating the limits of artificial intelligence without tool use. Achieving a 55% accuracy rate at the “high” reasoning setting would position Gemini Carbon well ahead of many current top-tier models, underscoring Google's accelerating progress in the race toward advanced general intelligence. Experts note that surpassing the 55% threshold on such a demanding exam requires substantial improvements in complex problem-solving and abstract reasoning.

As the AI community awaits the formal addition of the Carbon model to the official HLE-Diamond results chart, expectations are mounting that the checkpoint will solidify Google's dominance in the frontier model landscape. The forthcoming score will serve as a definitive measure of the architectural advancements distinguishing Carbon from its predecessors and contemporaries.