Google DeepMind has added a dial that controls how much a model “thinks” when it responds to a question.

The latest update to the top Gemini AI model from Google DeepMind includes a dial for controlling how much the system thinks through a response. The new feature was designed to save developers money, but it also acknowledges a problem. Reasoning models are prone to overthinking and wasting money and energy.

There are a few tried and true methods to make an AI more powerful. The first was to increase its size by using more data for training, and the second was to give better feedback on what constitutes an appropriate answer. Google DeepMind, along with other AI companies, began to use a third approach toward the end last year: reasoning.

Jack Rae, principal research scientist at DeepMind, says, “We’ve really pushed on ‘thinking’.” These models, which are designed to solve problems logically, and spend more time finding an answer, gained prominence earlier this year, with the launch the DeepSeek R1 Model. AI companies are interested in them because they can improve an existing model by teaching it to be more pragmatic. The companies can avoid building a model from scratch.

The AI model costs more when it spends more time and energy on a query. Leaderboards for reasoning models show that a single task can cost up to $200. This extra time and money will help reasoning models better handle challenging tasks like analyzing code, or gathering information from many documents.

According to Koray Kavukcuoglu of Google DeepMind, chief technical officer, “the more you iterate on certain hypotheses and ideas, the more likely “it is to find the right thing.” This isn’t always true, however. Tulsee Doshi is the leader of the Gemini product team. She says that the model “overthinks” and specifically refers to Gemini Flash 2.5. This model includes a slider to allow developers to reduce how much it thinks. “For simple prompts the model does overthink.”

When the model spends more time on a problem than is necessary, it increases the cost of running the model for developers, and it worsens AI’s environmental footprint. Nathan Habib, a Hugging Face engineer who has studied overthinking in reasoning models, said that it is common. Habib says that in the rush to demonstrate smarter AI, many companies reach for reasoning models even when there is no nail visible. OpenAI said that when it announced a new nonreasoning model in February it would be its last. Habib says that the performance gain is “undeniable”but not for most tasks where AI is used. Even when reasoning is used to solve the right problem, it can go wrong. Habib showed me a leading reasoning system that was asked to solve an organic chemistry question. The model started off well, but about halfway through the reasoning process, its responses began to resemble a meltdown. It sputtered hundreds of times “Wait but …”. It took far longer to complete the task than a model that didn’t reason. Kate Olszewska who evaluates Gemini models for DeepMind says Google’s models are also prone to getting stuck in loops.

Google’s new “reasoning dial” is an attempt to solve this problem. It’s not built for the consumer version yet, but for app developers. Developers can specify how much computing power should be spent on a particular problem. The idea is to reduce the dial if a task doesn’t require much reasoning. When reasoning is enabled, the outputs of the model cost about six times as much to produce.

This flexibility is also due to the fact that it is not yet known when more reasoning will lead to a better result. Rae says, “It is really hard to define a boundary, like what is the best task for thinking right now.”

Some obvious tasks include coding, (developers may paste hundreds of lines into the model before asking for help), and generating expert level research reports. Developers might find it worthwhile to turn the dial up for these tasks. Developers will need to test and provide feedback more often to determine when low or medium settings are sufficient.

Habib believes that the amount of money invested in reasoning models is an indication that the old paradigm on how to improve models is changing. He says that “scaling laws are being changed.”

Companies are betting on longer thinking times over larger models to get the best results. Since several years, it’s been obvious that AI companies spend more money on inferencing – when models are “pinged”to generate an answer to something – than on training. This spending will accelerate once reasoning models take off. Inferencing also accounts for a growing percentage of emissions.

On the subject of AI models that “reason”, “think”or “think about”an AI model can’t perform these acts the way we use such words to describe humans. I asked Rae to explain why the company uses such anthropomorphic language. Kavukcuoglu claims that Google does not try to mimic any specific human cognitive process.

Google DeepMind may not be the only game in town, even if reasoning models continue dominating. The results of DeepSeek were released in December and early January. This sparked a drop in the stock market of nearly $1 trillion because it was marketed as a way to get powerful reasoning models for incredibly low prices. The model is called “open weight”which means that its internal settings (called weights) are made public, allowing developers run it themselves instead of paying to access proprietary Google or OpenAI models. (The term “open-source” is reserved for those models that reveal the data on which they were trained.)

Why use proprietary models when open ones, like DeepSeek, perform so well? Kavukcuoglu believes that in coding, math and finance, “there is a high expectation that the model be very accurate, precise, and be able understand really complex situations.” He expects that models that can deliver on this, whether they are open or closed, will win out. DeepMind believes that this reasoning will form the basis of future AI models which act on your behalf to solve problems. He says that reasoning is the key to building intelligence. “The moment a model starts to think, it has begun to act.”

The story was updated in order for “overthinking.”