Google’s Gemini 3 achieves top LMArena scores, beating GPT-5 Pro in reasoning and multimodal tasks.
Google’s latest foundation model, Gemini 3, has set a new state-of-the-art in large language model benchmarks, surpassing previous leaders including OpenAI’s GPT-5 Pro. Gemini 3 achieved an Elo score of 1501 on the LMArena leaderboard, the first to cross this critical threshold, and scored 37.4 on the Humanity’s Last Exam benchmark, a widely recognized test measuring reasoning and general intelligence in AI.
Notably, Gemini 3 supports advanced multimodal capabilities, enabling it to process and integrate text with other data types more effectively than prior models. This has garnered positive reviews and strong developer interest, positioning it as a formidable competitor in today’s rapidly evolving AI landscape. The model’s superiority in reasoning and benchmarks reflects Google’s sustained investment in foundational AI research. It also impacts commercial AI offerings, as seen with Meta testing Google’s TPU chips for enhanced performance and capacity, indicating broad industry movement towards leveraging such advanced models and infrastructure.[1][2]
Source: google.com