Google says think deeper, not longer, and cuts inference costs by half

For the past few years, the large language model industry has followed an unwritten rule: if you want a more accurate answer to a hard question, make the model think longer. Chain-of-Thought, in which a model builds a step-by-step line of reasoning before its final answer, became the gold standard. Developers stretched these chains, enlarged context windows and spent ever more compute on every query. But a new study by the University of Virginia and Google argues that all this time we have been mistaking verbosity for intelligence.

The idea behind the work is deceptively simple. The researchers asked whether every extra token in a reasoning chain really brings the model closer to the right answer. Or is a significant share of those tokens informational noise, repetition and treading water? To find out, the team introduced a new metric, Deep-Thinking Ratio. Instead of measuring reasoning length in tokens, the metric assesses what share of the reasoning contains productive logical steps, meaning those that lead to a solution rather than merely filling space.

The results were telling. The analysis showed that in the typical long reasoning chains of modern LLMs, a huge portion of intermediate steps carries no real semantic weight. A model may rephrase the same thought dozens of times, return to stages it has already passed and generate redundant explanations, and all of this costs real money. Every unnecessary token at inference time means GPU time, electricity and latency for the end user. For large services handling billions of requests, the sums are colossal.

The study's key achievement is that, by optimizing the reasoning process with Deep-Thinking Ratio in mind, the researchers achieved two things usually considered mutually exclusive. Model accuracy rose, because cutting unproductive steps effectively lowers the chance that the model gets lost in its own reasoning and reaches a wrong conclusion. And overall inference costs fell by roughly half, because the model generates far fewer tokens per query. This is not a trade-off between quality and cost, but a rare case where optimizing one parameter improves both.

To grasp the scale of this finding, it helps to recall the context. Inference cost is one of the industry's main headaches. OpenAI, Google, Anthropic and other companies spend billions of dollars on compute infrastructure, and a significant share of that goes specifically to generating answers for users. Models such as OpenAI's o1 and o3, as well as Google's Gemini with extended thinking, were designed specifically for long reasoning chains. If it turns out that half of that reasoning can be cut without harm, or more precisely, that the model can be taught not to generate it in the first place, the economic effect would be measured in hundreds of millions of dollars a year.

There is also a deeper theoretical angle. The study effectively questions the very paradigm of inference scaling that dominated 2024-2025. If "thinking longer" does not equal "thinking better", then the race to enlarge context windows and grow the compute budget for reasoning is a dead end. Instead, the industry may need to focus on the quality of each reasoning step rather than their number. This echoes how human thinking works: an expert solves a problem not because they think longer than a novice, but because each of their mental steps is more purposeful.

The practical consequences for developers and users could appear fairly quickly. Deep-Thinking Ratio is a metric that is relatively easy to integrate into existing model training and evaluation pipelines. Large labs can be expected to start using similar approaches in fine-tuning, and cloud providers in optimizing the cost of API calls. For end users, this means faster and still more accurate answers at the same or a lower price.

The study by Google and the University of Virginia reminds the industry of a truth that is easy to forget in the pursuit of scale: efficiency is not about "more" but about "more precise". The models of the future will probably not be those that think the longest, but those that know how to think to the point.