Gemini 2.5 Preview for Code

Writing code with these models works like handing a junior engineer a giant repo and a spec, then asking for a patch. Google pushed an updated Gemini 2.5 Pro Preview on May 6, 2025, with the pitch that it writes, edits, and transforms code better than the prior version. (blog.google) Google called the release “Gemini 2.5 Pro Preview (I/O edition)” and said it shipped a couple of weeks before Google I/O so developers could start using it sooner. The company aimed the update at front-end work, code editing, code transformation, and “agentic workflows,” where the model chains steps and tools to finish a job. (developers.googleblog.com) Google said the model moved to No. 1 on the WebDev Arena leaderboard, a benchmark based on human preferences for how well models build working, polished web apps. The company said the update beat the previous Gemini 2.5 Pro by 147 Elo points on that leaderboard. (blog.google) The model page says Gemini 2.5 Pro can take in up to 1,048,576 input tokens, roughly the size of a very large codebase or document set, and return up to 65,536 output tokens. Google lists support for text, code, images, audio, video, and Portable Document Format files, plus function calling, code execution, search grounding, and structured outputs. (ai.google.dev) That matters for coding tools because the hard part is often not writing one function. It is keeping track of thousands of files, understanding a design mockup, reading logs, calling tools, and then changing the right code without breaking the rest of the app. (docs.cloud.google.com) Google’s developer blog framed the update around building interactive web apps from prompts and around turning other media into software. One example showed the model creating a learning app from a single YouTube video, which Google tied to an 84.8% score on the VideoMME video-understanding benchmark. (developers.googleblog.com) Outside Google’s own claims, partner quotes in the launch posts pointed to a narrower selling point: fewer mistakes in real developer workflows. Cursor Chief Executive Officer Michael Truell said his team saw “a significant reduction” in failed tool calls, while Replit President Michele Catasta said the model performed well on latency-sensitive tasks. (blog.google, developers.googleblog.com) The release also fit Google’s broader model strategy at the time. DeepMind’s Gemini 2.5 technical report described Gemini 2.5 Pro as the company’s top reasoning model, built for coding, multimodal input, long context, and “agentic” systems that can plan and use tools across steps. (storage.googleapis.com) By April 2026, Google’s current developer docs listed Gemini 2.5 Pro as a stable model last updated in June 2025, while the Gemini API changelog had already moved on to the Gemini 3.1 series. That leaves the May 2025 preview as a snapshot of a specific race inside artificial intelligence coding tools: less about a chatbot demo, and more about shaving time off everyday software work. (ai.google.dev, ai.google.dev)