Gemini 2.5's Impact on Agentic Workflows

The Gemini 2.X series is engineered to be natively multimodal, supporting input from various sources like text, audio, images, video, and code repositories^[1]. The models can process long context inputs exceeding 1 million tokens, allowing them to comprehend vast datasets and handle complex problems from different information sources^[1]. Native tool use support further enhances these capabilities, enabling the models to interact with external tools and services to solve complex tasks^[1].

Gemini 2.5 Pro stands out as the most capable model in the Gemini 2.X family, exhibiting strong reasoning and coding capabilities^[1]. It excels at producing interactive web applications and demonstrates an ability for codebase-level understanding^[1]. Additionally, it showcases emergent multimodal coding abilities, making it suitable for complex agentic tasks^[1]. One notable feature is its capacity to process up to 3 hours of video content, demonstrating its enhanced multimodal understanding^[1].

While Gemini 2.5 Pro offers top-tier performance, Gemini 2.5 Flash provides excellent reasoning abilities at a fraction of the compute and latency requirements^[1]. This hybrid reasoning model is useful for complex tasks, offering a controllable thinking budget to balance quality, cost, and latency^[1].

Past Gemini models produced answers immediately following a user query, which constrained the amount of inference-time compute (Thinking) that the models could spend reasoning over a problem^[1]. Gemini Thinking models are trained with Reinforcement Learning to use additional compute at inference time to arrive at more accurate answers^[1]. These models can spend tens of thousands of forward passes during a “thinking” stage before responding to a question or query^[1].

Thinking is integrated with other Gemini capabilities, such as native multimodal inputs (images, text, video, audio) and long context (1M+ tokens)^[1]. The model can decide for itself how long to think before providing an answer for any of these capabilities^[1]. Users also have the ability to set a Thinking budget, which constrains the model to respond within a desired number of tokens, allowing them to trade off performance with cost^[1].

Gemini Deep Research is an agent built on top of the Gemini 2.5 Pro model, designed to strategically browse the web and provide informed answers to even the most niche user queries^[1]. The agent is optimized to perform task prioritization and identify when it reaches a dead-end while browsing^[1]. Since its initial launch in December 2024, the capabilities of Gemini Deep Research have improved, as evidenced by its performance on the Humanity’s Last Exam benchmark^[1]: scoring 26.9% in June 2025, or 32.4% with higher compute^[1].

Gemini 2.0 and 2.5 represent a shift towards delivering tangible real-world value, empowering users to address practical challenges within complex software environments^[1]. Pre-training efforts have focused on incorporating a greater volume and diversity of code data from repository and web sources^[1]. Post-training involved developing novel techniques incorporating reasoning capabilities and curating a diverse set of engineering tasks^[1]. These advancements are demonstrated in IDE functionalities, code agent use cases for complex operations within full repositories, and multimodal interactive scenarios such as end-to-end web and mobile application development^[1].

Ensuring the factuality of model responses to information-seeking prompts remains a core pillar of Gemini model development^[1]. Gemini 2.0 marked a leap as the first model family trained to natively call tools like Google Search, enabling it to formulate precise queries and synthesize fresh information with sources^[1]. Gemini 2.5 integrates advanced reasoning, allowing it to interleave search capabilities with internal thought processes to answer complex, multi-hop queries and execute long-horizon tasks^[1].

The models can also connect to tools like Google Search and code execution, utilizing multimodal input and a 1 million-token context length^[1]. The goal was to provide an economical model class which provides ultra-low-latency capabilities and high throughput per dollar^[1]. To advance Gemini’s capabilities towards solving hard reasoning problems, a novel reasoning approach, called Deep Think, blends in parallel thinking techniques during response generation^[1].

Gemini 2.5 Pro excels at transforming diverse inputs into interactive and functional applications, capable to take a PDF script of a play and generate a tool that allows drama students to practise their lines^[1]. It excels at a wide range of complex tasks, from those relevant for education to creative expression^[1].

Gemini is incorporated into a wide variety of Google products, including AI Overviews and AI Mode within Google Search, Project Astra, the audiovisual-to-audio dialog agent, Gemini Deep Research, NotebookLM, Project Mariner, and Google’s coding agent, Jules^[1].