Archive — US — techcrunch
‘Deep Think’ boosts the performance of Google’s flagship Google Gemini AI model
Google has introduced 'Deep Think', an enhanced reasoning mode for its Gemini 2.5 Pro AI model, boosting performance on various benchmarks by allowing multi-answer consideration. Announced at Google I/O 2025, it outperformed competitors in coding evaluations. Additionally, updates to the Gemini 2.5 Flash model improve efficiency and task performance, while a new Gemini Diffusion model offers faster outputs.
Key points
- Google announced the Deep Think mode enhancing the Gemini 2.5 Pro AI model during Google I/O 2025.
- Deep Think improves reasoning by allowing the model to contemplate multiple answers before responding.
- Performance improvements enabled Gemini 2.5 Pro to excel at LiveCodeBench and outperform OpenAI’s o3 in tests for perception and reasoning.
- Deep Think is currently available to trusted testers via the Gemini API, with safety evaluations ongoing before wider release.
- An updated Gemini 2.5 Flash model was launched, enhancing coding, multimodality, reasoning, and long context efficiency.
- Google introduced the Gemini Diffusion model, noted for its speed, providing outputs 4-5 times quicker than similar models.
Why it matters
- The introduction of Deep Think aligns with current AI trends focusing on enhancing model reasoning capabilities.
- Google's advancements are part of a broader competition in AI, particularly against rivals like OpenAI, emphasizing performance benchmarks.
- The push for trusted testing reflects the industry’s increasing focus on safety and ethical AI deployment.
- Improvements in Gemini models are significant as they showcase the evolution of AI applications in programming and data interpretation, crucial for future tech.
Google's innovations with Deep Think and updated Gemini models signify a pivotal advancement in AI, enhancing reasoning capability and performance while addressing safety, reflecting ongoing trends shaping the AI landscape.