Google Updates Gemini 2.5 Pro AI for Improved Coding Skills

Google has announced an update to its Gemini 2.5 Pro model, which the company claims significantly enhances its performance on various coding, reasoning, and knowledge benchmarks. The update, labeled as an "updated preview," builds upon the improvements introduced about a month ago and will be rolled out for general availability in a few weeks. The updated model is now available on Google's AI developer platforms, including AI Studio, Vertex AI, and the Gemini app. Performance Enhancements According to Google, Gemini 2.5 Pro continues to excel in programming tasks, notably leading on difficult coding benchmarks. The model demonstrates top-tier performance in highly challenging tests that assess reasoning, mathematics, science, and general knowledge. For instance, it scored 86.4% on the single-attempt GPQA diamond test, which evaluates science knowledge, and 88.0% on the single-attempt AIME 2025 mathematics benchmark. In coding, it achieved 69.0% on LiveCodeBench for code generation and 82.2% on Aider Polyglot for code editing. These scores are comparable or even superior to those of competing models from OpenAI, Anthropic, and Grok. Addressing User Feedback In response to user feedback from the previous 2.5 Pro release, Google has made improvements to the model's style and structure. The updated version is designed to provide more creative and better-formatted responses, enhancing its usability for developers. These enhancements aim to make the model more versatile and user-friendly, aligning with Google's goal of creating a more intelligent and adaptable AI tool. Model Capabilities Gemini 2.5 Pro supports multiple data types for inputs, including text, images, video, and audio, and can handle up to 1 million input tokens. However, its output is limited to text with a maximum of 64,000 tokens. The model's knowledge cutoff is set at January 2025, ensuring it stays current with the latest information and technologies. Additionally, it incorporates advanced features such as function calling, structured output, search as a tool, and code execution, making it suitable for a wide range of applications. Availability and Deployment Developers can immediately start building with the updated preview of Gemini 2.5 Pro through the Gemini API, accessible via Google AI Studio and Vertex AI. These platforms offer enhanced control over cost and latency through the introduction of "thinking budgets." The model is also being rolled out today in the Gemini app, providing a convenient platform for users to interact with and test the new capabilities. Key Benchmarks LiveCodeBench (Code Generation): Single attempt score of 69.0%, showing strong performance in generating functional code. Aider Polyglot (Code Editing): Scores ranged from 79.6% to 72.0% in different scenarios, indicating high proficiency in editing and refining existing code. SWE-bench Verified (Agentic Coding): Single attempt score of 59.6% and multiple attempts score of 67.2%, demonstrating capability in collaborative coding tasks. GPQA diamond (Science Knowledge): Single attempt score of 86.4%, reflecting advanced scientific reasoning skills. Humanity's Last Exam (HLE): Scored 21.6% on a no-tools version, showcasing its reasoning and knowledge capabilities. AIME 2025 (Mathematics): Single attempt score of 88.0% and multiple attempts score of 88.8%, highlighting strong mathematical problem-solving abilities. Cost and Pricing The pricing for using Gemini 2.5 Pro varies based on the type and volume of inputs and outputs. For input prices, the cost is $1.25 per 1 million tokens without caching, while output prices stand at $10.00 per 1 million tokens. Compared to competitors like OpenAI and Anthropic, Google's model offers competitive pricing and flexibility. Multilingual Support The updated model performs exceptionally well across multiple languages, achieving an 89.2% score on the Global MMLU (Lite) benchmark. This makes Gemini 2.5 Pro a robust choice for international teams and multilingual projects. Industry Insights and Company Profile Industry experts view the updated Gemini 2.5 Pro as a significant step forward in AI development, particularly in the realm of coding and reasoning. The model's ability to handle a wide variety of data types and its impressive performance on multiple benchmarks solidify its position as one of the most advanced AI tools available. Google's commitment to addressing user feedback and continuously refining its models underscores the company's dedication to innovation and user satisfaction. With a robust feature set and competitive pricing, Gemini 2.5 Pro is well-positioned to become a preferred choice for developers working on complex and demanding projects.