Gemini 2.5 Pro continues dominating coding benchmarks, leading WebDev Arena with an ELO score of 1415 and outperforming rivals in LMArena’s human preference rankings. Google DeepMind’s experimental Deep Think mode, now integrated into 2.5 Pro, tackles complex math and coding tasks using multi-hypothesis reasoning, achieving 84.0% on MMMU multimodal tests. These updates build on the model’s 1 million-token context window and LearnLM integration, which educators praised for superior pedagogy in learning scenarios.

2.5 Flash gains efficiency improvements, reducing token usage by 20-30% while enhancing reasoning, code generation, and long-context handling. The model now powers the Gemini app for all users, with general availability in Google AI Studio and Vertex AI slated for early June. Its lightweight design makes it ideal for real-time applications, balancing speed and accuracy for developers.

New capabilities include native audio output for expressive dialogue, Project Mariner’s computer use tools for automating workflows, and advanced security against indirect prompt injections. Developers gain thought summaries in API responses, offering transparency into model decision-making, and MCP tool support for seamless open-source integrations. Text-to-speech now supports dual-voice outputs across 24 languages.

These advancements stem from cross-Google team efforts to refine safety protocols and expand tool accessibility. While Deep Think remains in testing with trusted users, broader rollouts prioritize responsible deployment. The updates underscore Google’s focus on practical AI applications, from coding assistants to multilingual conversational agents, positioning Gemini 2.5 as a versatile leader in both research and production environments.