Google’s Gemini 2.5 Pro AI Model Launched; Tops Leaderboard, Outperforms OpenAI’s o3 Mini

Google released the successor to its Gemini 2.0 series artificial intelligence (AI) models on Wednesday. Dubbed Gemini 2.5 Pro Experimental, it is the first model the company is releasing from the 2.5 family. The Mountain View-based tech giant says that this series of models will have “thinking” or reasoning capability built directly into the models. It also notes improved benchmark scores across a wide range of functions, outperforming OpenAI's o3-mini in several areas. Google has begun rolling out the model to users.

Gemini 2.5 Pro AI Model Released

In a blog post, Koray Kavukcuoglu, the CTO of Google DeepMind, detailed the new large language model (LLM). The most notable aspect of the Gemini 2.5 series is that there will no longer be any “Thinking” models such as the Gemini 2.0 Flash Thinking.

The tech giant used an enhanced base model, which was further improved in post-training to deliver inherent reasoning capabilities to all the Gemini 2.5 AI models. Hence, Google will not denote a particular “Thinking” label to a model as all of them can carry out advanced reasoning and show chain-of-thought (CoT).

Gemini 2.5 Pro benchmarks
Photo Credit: Google

Google did not reveal a lot about the model specifications, so details around its dataset, training methods, and architecture are not known. However, the tech giant shared its benchmark scores based on internal testing. It is said to have scored 18.8 percent on Humanity's Last Exam, a dataset considered the toughest benchmarking test for AI models. Gemini 2.5 Pro's score was state-of-the-art (SOTA) among models without tool use.

Gemini 2.5 Pro is also claimed to have outperformed models like OpenAI's o3-mini, Grok 3 Beta, Claude 3.7 Sonnet, and DeepSeek R1 in several benchmarks, such as GPQA Diamond, AIME 2024 and 2025, Aider Polyglot, and MMMU.

Apart from this, Gemini 2.5 Pro also ranked at the top of the LMArena leaderboard at release. LMArena is a user-based platform where AI enthusiasts and developers rate models based on their experiences. Currently, it is followed by Grok 3 preview, GPT 4.5 preview, Gemini 2.0 Flash Thinking, and Gemini 2.0 Pro for the second, third, fourth, and fifth positions, respectively.

Google claims that the latest LLM also improves coding performance and can create “visually compelling” web apps and agentic code applications. Gemini 2.5 Pro also comes with native multimodal support and a context window of one million tokens.

Gemini 2.5 Pro is available to developers and enterprises via the Google AI Studio, and Gemini Advanced subscribers can access the model in Gemini's web client and apps. The company plans to make it available on Vertex AI in the coming weeks.