Google Gemini 2.5 AI model beats GPT-4.5 in coding and reasoning | Features, access, performance

Google has officially announced Gemini 2.5, calling it their “most intelligent AI model yet”. Built to handle complex reasoning and coding tasks, Gemini 2.5 is the latest upgrade in the company’s multimodal AI lineup and arrives as an experimental version of Gemini 2.5 Pro. This version is already topping benchmark charts, including LMArena, and is now available to developers and advanced users.

While most AI models today still focus on quick responses or summarised outputs, Google says Gemini 2.5 is different. It thinks. Literally. As Koray Kavukcuoglu, CTO of Google DeepMind, puts it, “Gemini 2.5 models are thinking models, capable of reasoning through their thoughts before responding, resulting in enhanced performance and improved accuracy.”

Gemini 2.5: Built for smarter reasoning

The Gemini 2.5 Pro Experimental model is aimed at developers and professionals who need deeper reasoning and higher-quality outputs. The model scores the highest in math and science benchmarks like GPQA and AIME 2025 without relying on test-time tricks like majority voting. It also posts a standout 18.8% score on Humanity’s Last Exam, which was designed to test reasoning at the edge of human knowledge.

This isn’t just about text. Gemini 2.5 Pro supports multimodal inputs—text, code, images, audio, video, and even large codebases. It ships with a massive 1 million token context window and Google says a 2 million token support is coming soon. That means the model can process large datasets in one go without splitting the input.

Gemini 2.5: Big leap in coding

Google says it’s also made major gains in coding. Gemini 2.5 Pro scores 63.8% on SWE-Bench Verified, a leading benchmark for code evaluation. It’s designed to generate web apps, agentic code, transform or edit existing code—and do it with better understanding of what the user is trying to build.

The AI model also understands how to break down logic better than its predecessors, making it ideal for building apps or games from simple prompts. According to Google, 2.5 Pro is able to create a functional game just from a one-line prompt—something that was not easy with older versions.

Gemini 2.5: Availability and what comes next

Developers can try Gemini 2.5 Pro now via Google AI Studio and the Gemini app for advanced users. It will roll out to Vertex AI soon, and pricing details are expected in the coming weeks.

Google also confirmed that all future models will come with these new “thinking” capabilities baked in.

“Going forward, we’re building these thinking capabilities directly into all of our models,” the company said in its blogpost, “so they can handle more complex problems and support even more capable, context-aware agents.”

As OpenAI, Anthropic, and DeepSeek roll out their own next-gen models, Gemini 2.5 gives Google a strong card in the rapidly heating AI race.