Google Unveils Experimental Gemini Pro 2.5, Its Most Advanced AI Yet
March 2025
Back to News

Google Unveils Experimental Gemini Pro 2.5, Its Most Advanced AI Yet

Google DeepMind has announced the release of Gemini Pro 2.5 Experimental.

Google DeepMind has announced the release of Gemini Pro 2.5 Experimental, marking a significant advancement in the company's artificial intelligence capabilities 1. This new model is being hailed as Google's most intelligent AI to date, designed with a focus on tackling increasingly complex problems 1. The announcement suggests a strong drive to position Gemini Pro 2.5 Experimental at the forefront of AI development, potentially surpassing competitors in specific areas. The emphasis on handling "increasingly complex problems" indicates that this model is engineered for advanced reasoning and problem-solving, moving beyond the capabilities of previous AI iterations 2.

A key feature of Gemini 2.5 Pro Experimental is its enhanced ability to "think" through prompts before generating a response 2. This involves a sophisticated process of analyzing information, drawing logical conclusions, incorporating context, and making informed decisions, leading to improved accuracy in its outputs 2. Beyond reasoning, the model demonstrates a substantial improvement in coding performance when compared to earlier Gemini models, particularly version 2.0 1. It exhibits proficiency in creating visually compelling web applications and agentic code applications, showcasing its potential for more intricate software development tasks 2. Furthermore, the model is capable of code transformation and editing, highlighting its utility in streamlining software development workflows 2. Quantifiable evidence of its coding prowess is seen in its performance on SWE-Bench Verified, an industry benchmark for evaluating agentic code capabilities, where it achieved a score of 63.8% with a custom agent setup 2. This focus on a more human-like approach to problem-solving in AI could lead to more dependable and nuanced results. The significant advancements in coding suggest Google's dedication to establishing Gemini as a leading tool for software development, directly challenging other prominent AI models in this field. The ability to create complex applications hints at the model's potential to automate more advanced and creative coding tasks.

Gemini 2.5 Pro Experimental incorporates native multimodality, allowing it to understand and process information from diverse sources such as text, audio, images, video, and even entire code repositories 2. This integrated approach to handling various data types could lead to richer and more context-aware interactions. The model is also equipped with a substantial 1 million token context window, with plans to expand this to 2 million tokens in the near future 1. This expansive context window enables the model to comprehend large datasets and manage intricate problems involving multiple information sources. Additionally, it features an output token limit of 64,000 tokens, allowing for the generation of detailed and comprehensive responses 7. This substantial context window provides a considerable advantage, allowing the model to process and retain information from lengthy documents or conversations, which is essential for complex tasks like code analysis or document summarization. The planned increase to a 2 million token context window indicates Google's ambition to further enhance contextual understanding.

The capabilities of Gemini 2.5 Pro Experimental are further highlighted by its performance on various industry benchmarks. It has secured the top position on the LMArena leaderboard, which evaluates AI models based on human preferences, suggesting that its outputs are not only powerful but also align well with user expectations for quality and style 1. On the challenging Humanity's Last Exam (without tools), the model achieved a state-of-the-art score of 18.8%, demonstrating its advanced reasoning and knowledge recall 1. In math and science, Gemini 2.5 Pro Experimental shows strong performance on benchmarks like GPQA and AIME 2025, achieving these results without relying on computationally intensive techniques like majority voting 2. Its score of 63.8% on SWE-Bench Verified further confirms its strong capabilities in handling complex software engineering tasks 2. Additionally, the model exhibits leading scores in long-context reading comprehension (MRCR) and multimodal understanding (MMMU) 2. This robust performance across a range of benchmarks indicates a versatile and highly capable AI model. The achievement of these results without relying on costly test-time techniques suggests an emphasis on efficiency and the model's inherent abilities.

Currently, Gemini 2.5 Pro Experimental is accessible to Gemini Advanced subscribers through the Gemini application and to developers via Google AI Studio 2. Plans are underway to make it available on Vertex AI in the near future, indicating an intention to cater to enterprise-level users 1. The primary target audience for this initial experimental release includes developers who can leverage its advanced coding and reasoning capabilities, Gemini Advanced subscribers who can benefit from its enhanced features, and enterprises looking for cutting-edge AI solutions 2. Google is currently offering free access to the model through Google AI Studio, encouraging experimentation and feedback from the developer community 2. This initial availability strategy suggests a phased rollout, starting with early access for paying subscribers and developers, followed by broader enterprise access. Offering free access to developers is a strategic move to encourage innovation and gather valuable feedback.

Looking ahead, Google intends to make Gemini 2.5 Pro available on Vertex AI in the coming weeks, further expanding its reach 1. The company will also be announcing pricing details for scaled production use, which will include higher rate limits 2. The planned expansion of the context window to 2 million tokens will further enhance the model's capabilities 2. Notably, Google plans to integrate the "thinking capabilities" demonstrated in Gemini 2.5 Pro into all future models, indicating a fundamental shift in their AI development strategy 2. The company is also actively seeking feedback from users to facilitate continuous improvement of the model 2. The introduction of pricing for higher usage on Vertex AI suggests a tiered access model for different user needs. The integration of advanced reasoning into all future models signifies a long-term commitment to enhancing AI intelligence across Google's products. The emphasis on user feedback highlights the experimental nature of this release and Google's dedication to iterative development.

In conclusion, the release of Gemini Pro 2.5 Experimental represents a significant step forward in artificial intelligence. Its enhanced reasoning and coding abilities, coupled with superior multimodal and contextual understanding, position it as a powerful tool with the potential to impact various fields. As the experimental phase progresses and Google incorporates user feedback, this advanced AI model is likely to pave the way for further innovations and wider adoption.

Works cited

  1. Google releases 'most intelligent' experimental Gemini 2.5 Pro ..., accessed March 30, 2025, https://www.zdnet.com/article/google-releases-most-intelligent-experimental-gemini-2-5-pro-heres-how-to-try-it/

  2. Gemini 2.5: Our newest Gemini model with thinking - The Keyword, accessed March 30, 2025, https://blog.google/technology/google-deepmind/gemini-model-thinking-updates-march-2025/

  3. Gemini 2.5 Pro is Google's 'most intelligent AI model' with thinking built-in - 9to5Google, accessed March 30, 2025, https://9to5google.com/2025/03/25/gemini-2-5-pro-experimental/

  4. Google's Gemini 2.5 Pro model tops LMArena by close to 40 points - R&D World, accessed March 30, 2025, https://www.rdworldonline.com/googles-gemini-2-5-pro-model-tops-lmarena-by-40-points-outperforms-competitors-in-scientific-reasoning/

  5. Google Workspace Updates Weekly Recap - March 28, 2025, accessed March 30, 2025, http://workspaceupdates.googleblog.com/2025/03/release-notes-03-28-2025.html

  6. Google Gemini 2.5 Pro: The Thinking AI That's Redefining ... - Medium, accessed March 30, 2025, https://medium.com/@cognidownunder/google-gemini-2-5-pro-the-thinking-ai-thats-redefining-intelligence-22ce8c545e9a

  7. Gemini 2.5 Pro: Features, Tests, Access, Benchmarks & More ..., accessed March 30, 2025, https://www.datacamp.com/blog/gemini-2-5-pro

  8. Gemini 2.5 Pro Experimental: Top AI Tool for Entrepreneurs in 2025 ..., accessed March 30, 2025, https://medium.com/the-ai-entrepreneurs/gemini-2-5-pro-experimental-top-ai-tool-for-entrepreneurs-in-2025-2c56b471352d

  9. Gemini 2.5 Pro Exp: How to Access, Features, Applications & More, accessed March 30, 2025, https://www.analyticsvidhya.com/blog/2025/03/gemini-2-5-pro-experimental/

  10. Gemini Pro - Google DeepMind, accessed March 30, 2025, https://deepmind.google/technologies/gemini/pro/

  11. Google Unveils Gemini 2.5 Pro - Perplexity, accessed March 30, 2025, https://www.perplexity.ai/page/google-unveils-gemini-2-5-pro-8y6HVgqERqazaRajoCx0cg

  12. Putting Gemini 2.5 Pro through its paces - Simon Willison's Weblog, accessed March 30, 2025, https://simonwillison.net/2025/Mar/25/gemini/

  13. Google Introduces Gemini 2.5 Pro with Improved Reasoning and Coding Capabilities - InfoQ, accessed March 30, 2025, https://www.infoq.com/news/2025/03/gemini-2-5-pro/

  14. Gemini models | Gemini API | Google AI for Developers, accessed March 30, 2025, https://ai.google.dev/gemini-api/docs/models

  15. Google Gemini 2.5 Pro: The best LLM ever | by Mehul Gupta | Data Science in your pocket, accessed March 30, 2025, https://medium.com/data-science-in-your-pocket/google-gemini-2-5-pro-the-best-llm-ever-172d0665336b