Google Deep Thinking Ratio: why you won't be able to do without it

one hand holds up an illustration of a brain with the word AI in the center.

There is news circulating in artificial intelligence circles in recent weeks that deserves attention even if you don’t have a technology background.

I learned about this by reading the article“A New Google AI Research Proposes Deep-Thinking Ratio to Improve LLM Accuracy While Cutting Total Inference Costs by Half” on the MarkTechPost website.

Google has unveiled new research that could significantly change the way AI works and, consequently, how much it costs and how accurate it is.

It is called Deep Thinking Ratio, and it is the result of a joint project between Google and the University of Virginia.

In this article I explain what it means, why it is different from what is there now, and most importantly what could change for your business in the next 2-3 years.

No technicalities. Just what you need to know so you won’t be caught unprepared.

The problem Google has decided to solve

To understand the Google Deep Thinking Ratio, one must first understand how an AI thinks today.

When you ask an artificial intelligence system to do something complex — analyze sales data, suggest a pricing strategy, respond to a customer with an articulated problem — the model does not respond right away. It first “thinks.”

It generates a long sequence of intermediate steps, almost like a written argument, and then comes to the conclusion.

This technique has been around for a few years and is called Chain-of-Thought: literally, “chain of thought.”

It has greatly improved the performance of AI models on difficult tasks.

The problem is that over time a bad habit has developed: models have learned to produce longer and longer reasoning, convinced that more text = more quality.

The paradox: the more AI “thinks,” the worse it responds

Google researchers have discovered something counterintuitive: there is a negative correlation between the length of an AI’s reasoning and its accuracy.

This is not a subjective impression-it is a statistical measurement. The longer the generated text, the higher the probability that the final answer is wrong.

Why does this happen. AI models, when they produce very long texts, tend to fall into loops: they rewrite the same things in different ways, they amplify their initial errors, they lose the thread.

It is the phenomenon of “overthinking”: thinking too much does not help, in fact it harms.
I would say that these things, many times, also happen in reality 😅

💡 More text does not mean more intelligence. It is like the contributor who fills ten pages of reports to say what he could have said in three lines: at some point, quantity becomes the enemy of clarity.

What is the Google Deep Thinking Ratio: the quality of thinking, not the quantity

The Google Deep Thinking Ratio (DTR) stems from a simple question: if length of reasoning is an unreliable indicator, how can we measure whether an AI is really “thinking well”?

The answer is to look inside the model, not outside.

A large AI model processes each word through dozens of internal layers.

For “easy” words-articles, connectives, routine sentences-the model stabilizes its response already in the first layers.

For “difficult” words-those that require true logic, calculation, complex reasoning-the answer keeps changing and refining down to the last layers.

The researchers call these “deep-thinking tokens”: the words on which the model really does deep cognitive work.

The Deep Thinking Ratio measures the percentage of these tokens in total reasoning.

A high DTR means that AI is addressing complexity with the right intensity. A low DTR means it is producing routine text-perhaps a lot, but of little value.

The metaphor of the craftsman

Think of a craftsman who makes precision mechanical components.

There are operations he does almost automatically, without much cognitive effort: tightening a bolt, following a standard procedure. Then there are the moments when he stops, measures twice, reasons, consults technical drawings: these are the ones that determine the quality of the finished part.

The DTR measures exactly that: how much of AI’s work is genuinely difficult and deep, versus how much is simply well-packaged routine.

Think@n: less cost, more precision

The discovery of DTR led to a practical technique called Think@n. It works like this.

Instead of generating one long reasoning, the system initiates several reasonings in parallel – like opening several hypotheses at once.

After only 50 initial words, it calculates the DTR of each. Reasonings with low DTR (unpromising roads) are dropped immediately. Only those with high DTR are completed.

The result, tested on very complex mathematical problems, is as follows: 2% higher accuracy than current systems, 49% lower cost.

Almost half the expense less, with better results. For those running AI systems at scale, this is not a technical detail – it is an economic revolution.

What specifically changes for Italian SMEs

Let’s get to the part that concerns you most closely.

This technology is still in the academic research stage, but the time frame for adoption in AI has shortened tremendously.

It is reasonable to expect that tools based on Deep Thinking Ratio will begin to appear in commercial products by 2027.

What does it mean in practice? Let’s try to imagine it through four areas.

Food: more accurate demand forecasts at lower cost

a person is arranging food items on a countertop

A food company in northern Italy uses an AI system to forecast weekly production volumes based on seasonality, active promotions, and market trends.

Today the system generates long and expensive reasoning for each forecast.

With DTR, the same result – indeed a better one – would come at half the computational cost. The savings translate into budgets to expand analysis: more products monitored, more markets considered.

Fashion: customization without waste

a fashion designer is arranging a garment on the mannequin

A fashion company with Italian manufacturing uses AI to personalize product recommendations on digital channels.

The current problem: Models generate redundant descriptions and suggestions, with costs rising as the catalog grows.

A DTR system could automatically select the “densest” responses of genuine intelligence, discarding superficial ones. Less computational waste, more relevant communication to the customer.

Finance: more reliable risk analysis

a manager is analyzing financial data on a monitor

An SME that manages investments or lines of credit uses AI tools to assess the risk profile of its customers.

Today these systems produce long reports, but often with circular reasoning.

DTR would allow higher quality analyses to be automatically identified, reducing the risk of decisions based on AI-generated “fluff.” More reliability, less manual verification costs.

Mechanical: more efficient predictive maintenance

a technician on a production line in the mechanical industry is working on the computer

A 35-employee mechanical engineering company installed sensors on key machinery.

An AI system analyzes data every hour to predict failures. The volume of data is enormous: thousands of analyses per day.

With current systems, the computational cost is proportional to the amount of reasoning produced.

With DTR, the system could process twice as much data with the same budget, or maintain the same volume while reducing spending by 50 percent.

💡 Regardless of the sector, the message is the same: More efficient AI means being able to do more with the same resources-or saving resources to invest them in people.

The realistic timeline: what to expect and when

Understanding when and how this technology will enter the market allows you to plan smarter. Here’s a realistic estimate, based on the adoption times we’ve seen with similar innovations in recent years.

By the end of 2026, DTR is likely to be integrated into the flagship models of the big players (Google, OpenAI, Anthropic).

Not as a separate product, but as an internal improvement: answers will become more accurate and costs will begin to fall, even without you having to do anything.

In 2027, we expect business-oriented AI platforms–the ones SMEs use for CRM, marketing automation, data analytics–to begin integrating DTR-based selection systems.
The results will be more reliable, the costs lower.

Between 2027 and 2028, those who have already built AI processes within their companies will be able to adopt these improvements without starting from scratch.

Those who have not yet started, however, will face a double investment: learning how AI works and getting up to speed on new versions at the same time.

The risk of waiting

There is a mistake that many entrepreneurs make: waiting until the technology is “mature” before starting.
This is understandable-no one wants to invest in something that could change within a year.

But in the case of AI, waiting has a hidden cost.

Companies starting today are building internal skills, operational habits, structured data.

When DTR comes to commercial products, they will already know how to exploit it. The waiting companies are starting from scratch in an already advanced market.

Efficiency that respects people

I want to close with a reflection that for us at Factory Communication is never marginal.

The Google Deep Thinking Ratio is not just a matter of economic efficiency.

An AI that thinks better-with less waste, with greater accuracy-is also an AI that consumes less energy.

Data centers that run these models have a significant environmental impact. Improving efficiency by 50 percent is not just a savings to the business-it is a step toward more sustainable technology.

And then there is the human dimension. More efficient AI frees up resources.

Those resources can be reinvested in people: training, time, quality of work.

The goal is not to have an AI that does everything by itself. It is to have an AI that supports your people, frees them from repetitive tasks, and gives them room to do the things that really matter.

If you can then carve out some time, a game of foosball never hurts 😅

The DTR reminds us that even AI, to function well, must learn to distinguish what is worth doing deeply from what can be handled lightly.

It is not unlike how a good professional works-and how a company that wants to grow over time without burning out its people should work.

If you found this article interesting, you can visit the section of our site about AI for Business

Google Deep Thinking Ratio: what changes for Italian SMEs in the next 3 years