Google has unveiled major updates to its Gemini AI platform at its annual I/O 2025 developer conference, showcasing advancements in reasoning capabilities and educational applications.
The centerpiece announcement was the introduction of Deep Think, an experimental enhanced reasoning mode for Gemini 2.5 Pro. This new capability uses research techniques that enable the model to consider multiple hypotheses before responding. According to Google, 2.5 Pro Deep Think achieves impressive scores on the 2025 USAMO, currently one of the hardest math benchmarks, and leads on LiveCodeBench, a difficult benchmark for competition-level coding.
Google also announced that it has infused LearnLM directly into Gemini 2.5, positioning it as "the world's leading model for learning." As detailed in Google's latest report, Gemini 2.5 Pro outperformed competitors on every category of learning science principles. The model now tops popular real-world leaderboards like WebDev Arena for coding and LMArena, which measures human preference for model responses. Thanks to integration with LearnLM, Google's family of models fine-tuned for education, 2.5 Pro is being touted as the leading model for learning, preferred by educators for its teaching effectiveness.
Google CEO Sundar Pichai highlighted the rapid progress of Gemini models, noting that "Gemini 2.5 Pro sweeps the LMArena leaderboard in all categories." The company also showcased Gemini 2.5 Flash, described as a "powerful and most efficient workhorse model" that has been "incredibly popular with developers who love its speed and low cost." The new 2.5 Flash improves across key benchmarks for reasoning, multimodality, code, and long context, ranking second only to 2.5 Pro on the LMArena leaderboard.
Gemini 2.5 Flash is now available to everyone in the Gemini app, with general availability in Google AI Studio for developers and Vertex AI for enterprises coming in early June. Gemini 2.5 Pro will follow shortly after. Google attributes this progress to "the relentless effort of teams across Google to improve our technologies, and develop and release them safely and responsibly."
Powering these advancements is Ironwood, Google's seventh-generation Tensor Processing Unit (TPU), which the company describes as its "most powerful, capable and energy efficient TPU yet." Ironwood is purpose-built to power thinking, inferential AI models at scale and has been supporting Google's most demanding AI training and serving workloads.
With these updates, Google continues to push the boundaries of AI reasoning and educational capabilities, setting new benchmarks for performance while emphasizing responsible development and deployment.