Google Unveils Next-Gen AI Reasoning Models

Google Introduces Gemini 2.5: A New Era of AI Reasoning

Google recently announced the release of Gemini 2.5, a novel family of AI models distinguished by their ability to pause and deliberate before providing responses.

Gemini 2.5 Pro: The Flagship Model

The initial offering within this new family is Gemini 2.5 Pro Experimental, a multimodal AI model designed for advanced reasoning. Google asserts this represents their most capable model to date.

Access to Gemini 2.5 Pro will be granted on Tuesday through Google AI Studio, the company’s developer platform. It will also be available to subscribers of Gemini Advanced, Google’s $20 monthly AI plan.

The Future of Google's AI Models

Google intends to integrate reasoning capabilities into all future AI model developments.

The Rise of AI Reasoning Models

Following OpenAI’s launch of o1, the first AI reasoning model in September 2024, a competitive drive emerged within the technology sector to replicate and surpass its functionality. Currently, Anthropic, DeepSeek, Google, and xAI all offer AI reasoning models.

These models utilize increased computational resources and processing time to verify information and logically analyze problems before generating an answer.

Benefits and Costs of Reasoning

The implementation of reasoning techniques has led to significant advancements in AI performance on tasks involving mathematics and coding. Many experts believe reasoning models are crucial for the development of AI agents – autonomous systems capable of performing tasks with minimal human oversight.

However, it’s important to note that these advanced models typically come with higher operational costs.

Gemini 2.5: Google's Most Significant Effort

While Google has previously experimented with “thinking” AI models, releasing a version of Gemini with this capability in December, Gemini 2.5 signifies their most dedicated attempt to outperform OpenAI’s “o” series.

Performance Benchmarks

Google claims that Gemini 2.5 Pro demonstrates superior performance compared to its previous AI models and several leading competitors across various benchmarks.

The model was specifically engineered to excel in the creation of visually engaging web applications and agentic coding applications.

Code Editing Performance

On the Aider Polyglot evaluation, which measures code editing proficiency, Gemini 2.5 Pro achieved a score of 68.6%, exceeding the performance of top AI models from OpenAI, Anthropic, and DeepSeek.

Software Development Abilities

However, on the SWE-bench Verified test, assessing software development skills, Gemini 2.5 Pro scored 63.8%. This outperformed OpenAI’s o3-mini and DeepSeek’s R1, but fell short of Anthropic’s Claude 3.7 Sonnet, which achieved a score of 70.3%.

Multimodal Reasoning Capabilities

Regarding Humanity’s Last Exam, a multimodal assessment encompassing thousands of crowdsourced questions from mathematics, humanities, and natural sciences, Google reports a score of 18.8% for Gemini 2.5 Pro, surpassing most competing flagship models.

Context Window Size

Initially, Gemini 2.5 Pro will be launched with a 1 million token context window. This allows the AI model to process approximately 750,000 words in a single input – exceeding the length of the entire “Lord of The Rings” series.

Furthermore, Google plans to expand the input capacity to 2 million tokens in the near future.

API Pricing Information

Google has not yet released API pricing details for Gemini 2.5 Pro, stating that this information will be shared in the coming weeks.