Google has made its latest AI model, Gemini 2.5 Flash, available to everyone through the Gemini app, marking a significant advancement in its AI capabilities while prioritizing security.
The new 2.5 Flash model demonstrates improved performance across reasoning, multimodality, code, and long context benchmarks while becoming more efficient, using 20-30% fewer tokens in evaluations. Google plans to make it generally available in Google AI Studio for developers and in Vertex AI for enterprises in early June, with the more powerful 2.5 Pro version following soon after.
A standout feature of the upcoming Gemini 2.5 Pro will be Deep Think, an experimental enhanced reasoning mode specifically designed for highly complex mathematics and coding tasks. This mode leverages new research techniques that enable the model to consider multiple hypotheses before responding, with Google reporting impressive scores on challenging benchmarks like the 2025 USAMO mathematics test.
Security has been a major focus in this release. Google DeepMind has implemented advanced safeguards against indirect prompt injection attacks—a cybersecurity challenge where malicious instructions are embedded in data that AI models retrieve. Through automated red teaming and adversarial fine-tuning, Google has significantly increased protection rates against these attacks during tool use, making Gemini 2.5 the company's most secure model family to date.
The update also brings Project Mariner's computer use capabilities to the Gemini API and Vertex AI. This feature, which allows AI to observe and interact with web browsers to complete tasks, opens new potential applications for companies. Several organizations, including Automation Anywhere, UiPath, and Browserbase, are already exploring these capabilities, with broader rollout for developer experimentation planned for this summer.
For developers, Google is introducing thought summaries in the Gemini API and Vertex AI, providing greater transparency into the model's reasoning process. The company is also extending thinking budgets to 2.5 Pro, giving developers more control over how much processing the model does before responding.