Google Introduces Gemini 3 Flash as Default AI Model in Gemini App

Google has unveiled Gemini 3 Flash, the latest addition to its AI model lineup, emphasizing speed and efficiency. This model is now the default in the Gemini app and AI Mode in Search, replacing the previous Gemini 2.5 Flash.

Gemini 3 Flash demonstrates significant performance improvements over its predecessor. In the Humanity’s Last Exam benchmark, designed to assess expertise across various domains, it achieved a score of 33.7% without tool use, surpassing Gemini 2.5 Flash's 11% and closely matching the newly released GPT-5.2's 34.5%.

The model excels in handling multimodal content, allowing users to upload videos, sketches, or audio recordings and receive insightful analyses or interactive responses. For example, users can submit a short video of a sports activity and request performance tips, or provide a sketch for the model to interpret.

For developers, Gemini 3 Flash is accessible through the Gemini API in Google AI Studio, Google Antigravity, Gemini CLI, Android Studio, Vertex AI, and Gemini Enterprise. Companies such as JetBrains, Figma, Cursor, Harvey, and Latitude are already integrating this model into their platforms.

Pricing for Gemini 3 Flash is set at $0.50 per 1 million input tokens and $3.00 per 1 million output tokens. Despite being slightly more expensive than its predecessor, Google asserts that the new model outperforms Gemini 2.5 Pro while operating three times faster. Additionally, it uses 30% fewer tokens on average for complex tasks, potentially leading to overall cost savings.

Tulsee Doshi, Senior Director and Head of Product for Gemini Models, described Gemini 3 Flash as a "workhorse model," emphasizing its cost-effectiveness and suitability for bulk tasks.

This release comes amid heightened competition in the AI sector. OpenAI recently launched GPT-5.2 following an internal "code red" directive to accelerate development in response to Google's advancements. GPT-5.2 offers enhanced general intelligence, improved coding capabilities, and better handling of long-context understanding.