Aug 12, 09:43 AM

Leaked Tests Suggest Google’s Gemini 3.0 Beats GPT-5

Publicly accessible code references and user reports indicate that Google is testing its next-generation large language model, Gemini 3.0 Pro. Mentions of a companion variant dubbed “Flash” were also spotted in the company’s command-line interface on GitHub, fuelling expectations that a formal launch is close. An unverified leak of the “Humanity’s Last Exam” benchmark circulated on 12 August assigns Gemini 3.0 a score of 32.4%, ahead of OpenAI’s recently released GPT-5 at 26.5% and xAI’s Grok 4 at 23.9%. While the figures have not been confirmed by Google, they suggest the model could narrow the performance gap with ChatGPT in advanced reasoning tasks. Despite the potential gains, ChatGPT continues to dominate user traffic. Data compiled by Similarweb for the same day show OpenAI’s service holding 78.5% of global generative-AI visits, compared with 8.7% for Gemini and 2.5% for Grok, underscoring the commercial stakes attached to incremental accuracy improvements. Competition across the sector remains intense. Anthropic, for example, has begun rolling out a chat-memory feature for Claude to subscribers on its Max, Team and Enterprise plans, mirroring functionality already offered by ChatGPT. Google has yet to announce an official release date or comment on the leaked benchmark results.