EXPLORINEER
en

Archive — US — techcrunch

Google unveils a next-gen family of AI reasoning models

Google has launched Gemini 2.5, a cutting-edge AI reasoning model that 'thinks' before responding, claiming it’s their most intelligent model. Available via Google AI Studio and Gemini Advanced subscribers, this model excels in coding and math tasks. Competing against OpenAI's models, Gemini 2.5 Pro boasts improved performance benchmarks, although reasoning models come with higher costs.

Key points

  • Google introduced Gemini 2.5, an AI model with built-in reasoning capabilities.
  • The model is accessible in Google AI Studio and through the premium Gemini Advanced plan.
  • Gemini 2.5 Pro is claimed by Google to outperform both its previous models and competitors' models on several tests.
  • On the Aider Polyglot evaluation for code editing, Gemini 2.5 Pro scored 68.6%, outperforming OpenAI and Anthropic models.
  • In SWE-bench Verified for software development, it scored 63.8%, underperforming Anthropic’s Claude 3.7 Sonnet.
  • Gemini 2.5 Pro's context window allows processing of 1 million tokens, equating to approximately 750,000 words.
  • Future updates will enable double the input length at 2 million tokens.

Why it matters

  • The launch reflects a broader trend in AI development where reasoning capabilities are becoming essential.
  • AI reasoning models have emerged since OpenAI's first launch in September 2024.
  • The ongoing competition among tech firms like Anthropic, DeepSeek, and xAI indicates a rapid evolution in AI technology, focusing on minimizing human intervention.
  • The need for more computing power and the consequent increase in operational costs highlight challenges in the scaling of AI technologies.

In summary, Google’s Gemini 2.5 marks a significant advancement in AI reasoning, showcasing the company's efforts to assert dominance in a competitive landscape while facing the challenges of increased costs and technological demands.