Gemini 2.5 Flash Preview (Sep '25) (Reasoning) is a high-efficiency multimodal model from Google, released on September 25, 2025. This iteration introduces native thinking capabilities, allowing the model to perform internal chain-of-thought reasoning before generating a final response. This reasoning process is designed to improve performance on complex logical tasks, including advanced mathematics, scientific problem-solving, and software engineering.
One of the model's primary features is a reasoning control mechanism, which enables developers to set a "thinking budget" of up to 24,576 tokens. This allow for granular management of the trade-off between deep internal processing and response speed. The model also features a 1-million-token context window and supports multimodal inputs such as text, image, audio, and video.
Compared to earlier versions, the September 2025 preview shows significant gains in agentic tool use, notably reaching 54% on the SWE-Bench Verified benchmark. It is also optimized for reduced verbosity and improved token efficiency, making it better suited for high-throughput applications that require concise and structured outputs, such as automated summaries or multi-step agentic workflows.