Google's latest artificial intelligence model, Gemini 3 Pro, has achieved a record score of 37.5 per cent on Humanity's Last Exam, a benchmark designed to identify artificial superintelligence. The test, created by AI safety researchers, assesses whether AI can reason at the frontier of academic knowledge.
According to Google, this score demonstrates PhD-level reasoning and places Gemini 3 Pro ahead of rival models from Anthropic, Meta, and OpenAI. The closest competitor is OpenAI's GPT-5 Pro, which scored 31.64 per cent on the same exam.
Google CEO Sundar Pichai described Gemini 3 as a 'massive jump in reasoning', capable of understanding depth and nuance. The model will be integrated into Google's Search AI Mode and the Gemini app, which has over 650 million monthly active users.
Google also unveiled Gemini 3 Deep Think, a more powerful version that scored 41 per cent on Humanity's Last Exam and achieved a record 45.1 per cent on the ARC-AGI-2 benchmark. However, Deep Think will not be publicly released until further safety checks are completed.
