Google has announced Gemini 2.5, describing it as their “most intelligent AI model” to date. The initial release is an experimental version of 2.5 Pro, which Google claims ranks first on the LMArena benchmark “by a significant margin.”
A Thinking Model
What sets Gemini 2.5 apart is its ability to “think” before responding. Unlike standard AI that immediately generates answers, this model pauses to reason through problems, resulting in more accurate responses.
“In the field of AI, a system’s capacity for ‘reasoning’ refers to more than just classification and prediction,” Google explains in its press release. “It refers to its ability to analyze information, draw logical conclusions, incorporate context and nuance, and make informed decisions.”
Performance Claims
Google reports that Gemini 2.5 Pro Experimental outperformed competing models on several benchmarks:
- On Humanity’s Last Exam (HLE), a particularly difficult test designed to avoid the problem of benchmark saturation, Gemini 2.5 scored 18.8%, compared to OpenAI’s o3 mini at 14% and Anthropic’s Claude 3.7 Sonnet at 8.9%.
- On SWE-bench Verified, which measures software development abilities, Gemini 2.5 Pro scored 63.8%, outperforming some competitors but falling behind Anthropic’s Claude 3.7 Sonnet, which scored 70.3%.
Similar Posts
Technical Capabilities
The model launches with a 1 million token context window, allowing it to process approximately 750,000 words at once—more text than the entire “Lord of the Rings” book series. Google says this capacity will soon double to 2 million tokens.
Google also highlights improvements in reasoning, multimodal capabilities (handling text, images, code, audio, and video), and “agentic capabilities,” suggesting the model could be used for more autonomous tasks.
Availability and Pricing
Gemini 2.5 Pro is available now for Gemini Advanced users through Google AI Studio and the Gemini app. Google says it will be “coming to Vertex AI soon” and will release pricing information “in the next few weeks.”
This release follows Google’s earlier experiment with reasoning models, Gemini 2.0 Flash Thinking, which was introduced in December. According to Google, all future models will have reasoning capabilities “baked in.”
The race for more capable AI reasoning models has intensified since OpenAI launched the first such model, o1, in September 2024. Currently, Anthropic, DeepSeek, Google, and xAI all offer AI models with reasoning abilities.
Frequently Asked Questions
Gemini 2.5 is Google’s newest AI model that’s designed to “think” before responding. Unlike previous models that generate answers immediately, Gemini 2.5 pauses to reason through problems, leading to more accurate responses. Google describes it as their “most intelligent AI model” yet, claiming it outperforms competitors on several key benchmarks.
According to Google, Gemini 2.5 Pro Experimental outperforms competing models on several benchmarks. On Humanity’s Last Exam (HLE), it scored 18.8%, compared to OpenAI’s o3 mini at 14% and Claude 3.7 Sonnet at 8.9%. For code editing, it reached 68.6% on Aider Polyglot, beating several competitors. However, on software development tests (SWE-bench Verified), it scored 63.8%, which is behind Claude 3.7 Sonnet’s 70.3%.
A “thinking model” means the AI doesn’t immediately generate a response. Instead, it takes time to analyze information, draw logical conclusions, and incorporate context before answering. This reasoning process helps the model solve complex problems more accurately and provide more thoughtful responses, similar to how humans might pause to think through a difficult question.
Gemini 2.5 Pro launches with a 1 million token context window, meaning it can process approximately 750,000 words at once. For perspective, that’s more text than the entire “Lord of the Rings” book series. Google has announced plans to double this capacity to 2 million tokens in the future, further enhancing the model’s ability to handle large amounts of information.
Currently, Gemini 2.5 Pro is available to Gemini Advanced users (Google’s $20/month AI subscription) through Google AI Studio and the Gemini app. Google has stated it will be “coming to Vertex AI soon.” As for pricing, Google hasn’t released specific details yet but plans to share pricing information “in the next few weeks.”
Gemini 2.5 is a multimodal AI model, meaning it can process and connect information from various sources including text, images, code, audio, and video. This versatility makes it suitable for a wide range of applications, from analyzing documents and generating creative content to coding tasks and providing detailed explanations based on multiple types of input.