Google is tackling a growing problem in AI: overthinking. The tech giant has unveiled a new “thinking budget” feature for its Gemini 2.5 Flash model that allows developers to control the amount of computational resources the AI uses for reasoning. This addresses the issue of powerful AI models using excessive processing power for simple tasks, leading to higher costs and environmental impact.
The new control provides a granular approach, with settings ranging from zero reasoning to 24,576 tokens of “thinking budget”, offering developers the flexibility to fine-tune the model’s processing based on the task at hand. Google’s documentation reveals that full reasoning can increase output costs sixfold compared to standard processing. Experts at Hugging Face and DeepMind confirm that AI models often get trapped in computational loops, wasting resources without improving results.
This feature marks a shift in AI development, moving away from simply scaling up model size and focusing on efficiency. By optimizing reasoning processes, Google aims to reduce the energy consumption of AI responses, which now contributes significantly to technology’s carbon footprint. Competitors are also exploring similar efficiency-focused approaches. However, Google DeepMind believes that proprietary models will retain an edge in specialized domains that require high precision.
Ultimately, the “thinking budget” allows organizations to make informed trade-offs between processing power and operational costs, making advanced AI capabilities more accessible and sustainable. Google asserts that Gemini 2.5 Flash delivers comparable results to other leading models at a lower cost and size, a value proposition enhanced by this new level of control.