In a move that underscores both the promise and pitfalls of reasoning-based artificial intelligence, Google DeepMind has introduced a new feature in its latest Gemini model: a dial that allows developers to adjust how much the AI “thinks” before generating a response.
The update, reported by MIT Technology Review, reflects a broader shift within the AI industry toward models designed for deeper, more logical processing—commonly referred to as “reasoning.” But it also acknowledges an emerging challenge: these models are increasingly prone to overthinking, often pouring excessive time, energy, and compute power into relatively simple tasks, driving up costs and environmental impact.
The Cost of Thinking Too Much
Since late 2023, companies like Google, OpenAI, and DeepSeek have turned their focus from simply scaling model size to enhancing their reasoning capabilities. The idea is to improve performance on complex problems—like writing code, conducting deep analysis, or interpreting vast documents—without needing to build entirely new models from scratch.
“We’ve been really pushing on ‘thinking,’” said Jack Rae, principal research scientist at DeepMind. But as MIT Technology Review highlights, reasoning doesn’t always yield better results—and it certainly doesn’t come cheap. For some tasks, the cost of running a reasoning model can exceed $200 per query, according to internal benchmarks.
Developers can now use the new slider to limit how much compute power—and thus, how much “thought”—a Gemini model expends. The company’s goal is to give developers greater control, especially for use cases where intensive reasoning is unnecessary.
“For simple prompts, the model does think more than it needs to,” Tulsee Doshi, who leads the Gemini product team, told MIT Technology Review. Left unchecked, the model’s tendency to overanalyze can lead to runaway costs and degraded performance. One internal example cited showed a reasoning model trapped in a mental loop, repeating “Wait, but…” hundreds of times during a chemistry problem.
Rethinking the Paradigm
The shift to reasoning marks a departure from the traditional “scaling laws” that guided the development of earlier AI systems—namely, bigger models trained on more data. Nathan Habib, an engineer at Hugging Face interviewed by MIT Technology Review, called reasoning the “new frontier,” but warned that companies are applying it indiscriminately: “They’re reaching for reasoning models like hammers even when there’s no nail in sight.”
In the case of Gemini Flash 2.5, Google has not yet rolled out the reasoning dial to consumer-facing versions of the model. Instead, it’s currently available for developers building apps and enterprise tools. That audience can now set specific budgets for how much computational power they want to allocate to a given task, potentially bringing more efficiency and predictability to their applications.
“We don’t yet know when more reasoning is actually better,” admitted Rae. While reasoning excels at complex, open-ended queries—like debugging code or drafting research reports—it may add needless latency to simpler prompts.
Environmental and Ethical Considerations
Beyond economic costs, reasoning-heavy models also carry a steep environmental price. Inference—the process of generating responses—has already surpassed training in terms of energy use and emissions for major AI companies. With models now encouraged to “think more,” that footprint is only expected to grow.
Still, Google maintains that reasoning is essential for future AI systems with agency and autonomy. “The moment the model starts thinking, the agency of the model has started,” said DeepMind CTO Koray Kavukcuoglu. However, he stressed that the company is not attempting to mimic human cognition. Anthropomorphic language like “thinking,” he said, is just a helpful shorthand.
A Competitive Landscape
Despite Google’s push, it isn’t the only player in the reasoning race. Open models like DeepSeek R1, which gained traction after release in late 2024, are performing competitively while being publicly accessible. The buzz around DeepSeek briefly wiped nearly $1 trillion from the tech-heavy stock market due to fears that open-weight models might rival proprietary offerings from Google and OpenAI.
Kavukcuoglu acknowledged the competition but argued that for domains like coding, math, and finance—where precision and reliability are paramount—premium models with robust reasoning will still be in high demand.
“The promise of reasoning is not just better performance,” he said. “It’s a foundation for intelligent systems that can act on your behalf and solve problems for you.”
As reasoning becomes the new standard in AI development, Google’s dial may be just the first of many tools aimed at balancing intelligence with practicality.
