Google DeepMind Unveiled Gemini 2.5 With Advanced Thinking Capabilities

Updated: March 30 2025 12:28
Google DeepMind has unveiled Gemini 2.5, heralded as their most intelligent AI model to date. This is what Google calls "thinking models" designed to tackle increasingly complex problems through enhanced reasoning capabilities.

Like DeepSeek R1, ny incorporating reasoning capabilities directly into the foundation of the model, Gemini 2.5 can analyze information more thoroughly, draw logical conclusions with greater accuracy, and make more informed decisions while considering context and nuance.

The first release in this new generation is Gemini 2.5 Pro Experimental, which has already claimed the top spot on the LMArena leaderboard by a significant margin. This benchmark measures human preferences for AI-generated content, suggesting that Gemini 2.5 Pro not only performs complex tasks effectively but does so with a high-quality style that resonates with users.



Without relying on expensive test-time techniques like majority voting, the model has established itself as state-of-the-art across a range of benchmarks that demand advanced reasoning skills.


In mathematics and scientific reasoning, Gemini 2.5 Pro leads in challenging tests like GPQA and AIME 2025. Perhaps most impressively, it scored 18.8% on Humanity's Last Exam – a dataset specifically designed by hundreds of subject matter experts to test the absolute frontier of human knowledge and reasoning ability.

The benchmark results suggest that Gemini 2.5 Pro isn't just faster or more efficient than its predecessors – it's fundamentally more intelligent in how it approaches complex problems.



For developers and software engineers, Gemini 2.5 Pro's coding abilities may be its most exciting feature. Google DeepMind has clearly prioritized coding performance in this release, achieving what they describe as "a big leap over 2.0." Gemini 2.5 Pro excels at:

• Creating visually compelling web applications
• Building agentic code applications
• Code transformation and editing
• Understanding and working with entire codebases

On SWE-Bench Verified, the industry standard for evaluating agentic code capabilities, Gemini 2.5 Pro achieved an impressive 63.8% score with a custom agent setup. This translates to real-world applications where the model can generate executable code from simple prompts – even creating functional video games from a single line prompt as shown below:



Gemini 2.5 model ships with a 1 million token context window, with plans to expand to 2 million tokens soon. This represents one of the longest context windows available in commercial AI systems today.

This expanded memory allows Gemini 2.5 Pro to comprehend vast datasets and handle complex problems that require integrating information from multiple sources. The model can seamlessly work with:

• Text documents
• Audio files
• Images
• Video content
• Code repositories

Researchers can analyze comprehensive datasets spanning different media types, developers can process entire codebases to understand complex software architecture, and creative professionals can generate content that integrates multiple formats.


According to the podcast updates, Google's internal testing process offers fascinating insights into how they evaluate their models. For "vibe checks," team members start with simple prompts like "Hey, how are you?" before moving to more complex requests like creating poems with specific constraints or building games.


For example, they mentioned spending "an unreasonable amount of time on Saturday playing Snake" because they had created a web app using Gemini 2.5 Pro from a single prompt. Another popular test is asking the model to create a simulation of a ball bouncing around a square, which tests both graphics generation and physics understanding.

The model's video understanding capabilities are particularly impressive. The team highlighted how Gemini 2.5 Pro combines:

• Strong multimodal understanding for visual content
• Long context handling for multi-hour videos
• Reasoning capabilities to analyze and identify key moments

One example mentioned was analyzing cricket matches (which can run for hours) to identify specific events like when a wicket is taken.


For those eager to experience Gemini 2.5's capabilities firsthand, Google has made the Pro version available through multiple channels:

• Available now in Google AI Studio for developers
• Accessible to Gemini Advanced users in the Gemini app (via model dropdown on desktop and mobile)
• Coming soon to Vertex AI for enterprise applications

Gemini 2.5 Pro is taking off 🚀🚀🚀

The team is sprinting, TPUs are running hot, and we want to get our most intelligent model into more people’s hands asap.

Which is why we decided to roll out Gemini 2.5 Pro (experimental) to all Gemini users, beginning today.

Try it at no… https://t.co/eqCJwwVhXJ

— Google Gemini App (@GeminiApp) March 29, 2025

The team outlined several upcoming developments:

• Pricing and full production release of Gemini 2.5 Pro (currently in experimental release)
• Bringing the 2.5 series capabilities to other models in the family, starting with Flash
• Optimizing the "thinking" process so models don't overthink simple prompts
• Providing developers more control over the thinking process for cost and latency optimization
• Image generation capabilities

Looking further ahead, the team mentioned continuing work on UI control and agent-like capabilities that were previously discussed with the 2.0 Flash release in December.

What's perhaps most telling is the accelerated pace of Gemini development - moving from 2.0 to 2.5 in just three months indicates that we're in a period of rapid innovation where capabilities are evolving faster than ever before.

Gemini 2.5: Google AI Studio
Gemini 2.5: Gemini App
Gemini 2.5: Vertex AI