Google Launches New Gemini 2.5 AI Model, Claims Superior Performance

Google has recently updated its AI models with the release of Gemini 2.5, following the launch of an "open" AI model optimized for single-GPU performance. This new version combines a significantly improved base model with enhanced post-training techniques to boost overall capabilities.

Gemini 2.5 Pro, now available in beta, surpasses competitors like OpenAI, Anthropic, xAI, and DeepSeek in common AI benchmarks, including comprehension, mathematical reasoning, and coding proficiency. The model is accessible via Google AI Studio or as an option in the model dropdown menu for Gemini Advanced subscribers.

In addition, Google highlighted the native multimodal capabilities of the Gemini series, which supports diverse data types such as text, audio, static images, video, and code. The company also announced plans to introduce a context window supporting up to 2 million tokens, further enhancing its data processing capacity.

Demis Hassabis, CEO of Google DeepMind, shared on social media that Gemini 2.5 Pro leads in LMArena with a 39-point higher ELO score. It demonstrates notable improvements in multimodal reasoning, coding, and STEM-related tasks.

According to Google, the leap in Gemini's quality stems from the integration of a "reasoning" mechanism. This allows the models to process tasks step-by-step and make more informed decisions, resulting in better responses to complex prompts. Google is embedding these reasoning abilities into all its models to handle more intricate problems and enable smarter, context-aware agents.