Cost Reduced to 1/76. What Would Happen If This Occurred in Your Company?

Asana has utilized GPT-6.1 to reduce the model cost of its browser agent to 1/76. At the same time, performance has increased by 5 times.

The security company Sophos has integrated OpenAI’s Daybreak to cut threat investigation time by 96%, automating 52% of MDR cases while still retaining human oversight.

On the other hand, OpenAI’s revenue forecast for this year has dropped to $5 billion, down from an initial estimate of $7 billion — a $2 billion shortfall. AI surveillance startup Flock has laid off about 270 employees, having been sidelined by privacy issues.

There are companies achieving dramatic results with AI, while others are sinking with the same technology. Where does this disparity come from?

To put it simply, it comes down to whether a company is deciding “What costs to eliminate with AI” rather than “What to do with AI”. That’s all there is to it.

—

Asana’s 1/76 — It’s Not Just About Changing the Model

At first glance, the Asana case seems to suggest that “costs went down simply by switching to a new model.” However, the essence is much deeper than that.

What Asana did was break down the cost structure of its browser agent — an AI function that operates tasks on behalf of users — and identify which layer of cost was the heaviest. They then redesigned the architecture to align with the characteristics of GPT-6.1 (low inference costs and efficient long-context handling).

In other words, analyzing the cost structure came first, followed by model selection. Many companies reverse this order. If they start with “Let’s try GPT-4o for now,” they proceed without a clear understanding of what costs they want to eliminate. As a result, they end up thinking, “We implemented AI, but it didn’t have the expected effect.”

Let’s consider the impact of the figure 1/76 concretely. Suppose the monthly model usage fee was 3 million yen. At 1/76, it would be about 40,000 yen. This means a savings of 2.96 million yen per month. Over a year, that amounts to over 35 million yen, equivalent to the salaries of 2-3 employees in a small to medium-sized enterprise.

Moreover, performance has increased fivefold. Costs have decreased while quality has improved. When such changes occur, it’s not just about “doing the same thing cheaper”; rather, a reversal happens where “previously unprofitable tasks become profitable.” Asana has achieved precisely that, enabling the deployment of high-functionality agents that were previously cost-prohibitive to all customers.

—

Sophos’s 96% Cut — Mastering the Art of Automation

The Sophos case is even more enlightening.

They have reduced the time for investigating cyber threats by 96%. What used to take 100 hours now takes just 4 hours. Moreover, they have automated 52% of MDR cases.

What’s noteworthy here is the “52%” figure. It’s not 100%.

Sophos has intentionally retained human oversight. In the realm of security, the risks associated with AI misjudgments are significant. Therefore, they designed the system so that “AI makes the initial judgment, and humans perform the final verification.” The goal was not to push automation rates close to 100%, but rather to create a state where “humans can concentrate on cases that truly require judgment.”

This structure applies equally to AI implementation in small to medium-sized enterprises. If one thinks, “I want to leave everything to AI,” they will likely fail. However, if they consider, “Let’s increase the density of tasks that humans should handle,” they will succeed.

With a 96% reduction in time, security analysts can now dedicate more time to advanced threat analysis and customer interactions. AI eliminates commodity tasks, thereby increasing the value density of human work. This is the essence of ROI.

—

OpenAI’s $2 Billion Shortfall, Flock’s 270 Layoffs — What Went Wrong?

Now, what about those who are struggling?

OpenAI’s revenue has fallen short of projections by $2 billion. This is not a reflection that “AI technology is flawed.” Rather, it highlights a structural issue: “Selling technology alone makes it difficult to achieve ROI.”

OpenAI is on the side providing APIs. Unlike Asana and Sophos, which integrate AI into their specific business challenges, they sell generic AI capabilities. Because it is generic, if customers cannot decide “What to use it for,” they do not reach a contract. Even if they do contract, they may not be able to utilize it effectively and end up canceling.

The Flock case is even clearer. They created a technically viable product in AI surveillance cameras, but misjudged whether society would accept it. Growing backlash over privacy issues halted contracts from municipalities and companies. What is technically “possible” does not equate to what can be “sold” in the market.

The common thread is that the question of “whose costs to reduce, and by how much” was ambiguous. The brilliance of the technology or the grandness of the vision does not yield ROI.

—

The Real Turning Point for Small and Medium-Sized Enterprises

Let’s bring this discussion closer to local small and medium-sized enterprises.

Asana is a tech company with a team of several hundred engineers. Sophos is also a global security firm. You might think, “Our scale is different.”

However, the structure is the same.

Even in small and medium-sized enterprises, changes close to the “1/76 cost reduction” are occurring. For instance, some companies have started using Claude or GPT to handle report creation tasks that previously cost 300,000 yen per month in outsourcing. The monthly API usage fee is just a few thousand yen. The costs have effectively dropped to less than 1/50.

A company that used to spend 3 hours a day responding to inquiries has automated initial responses with an AI chatbot, leaving only complex cases for human handling. Response time has been reduced by 70%. With the time saved, they have been able to focus on sales, resulting in an increase of 500,000 yen in monthly revenue.

Both cases indicate that while they have “implemented AI,” the reasons for their success are clear.

1. They identified the costs to eliminate before selecting AI (report outsourcing costs, inquiry response time).
2. They did not aim for 100% automation, leaving room for human value contribution.
3. They measured effectiveness numerically (from 300,000 yen to a few thousand yen, from 3 hours to 1 hour).

Conversely, the patterns of failure are also clear. They start with “I want to do something with AI.” They try out trendy tools within the company. It’s interesting but fails to establish itself in operations. Three months later, no one is using it. For companies where “using AI” becomes the goal, this path is almost inevitable.

—

So, What Should Be Done?

You only need to do three things.

1. List “What are you paying monthly?”

Outsourcing costs, labor costs (in time equivalents), software usage fees. First, put the numbers down. Without numbers, it’s unclear what can be eliminated with AI.

2. Start by eliminating “the most expensive and the most mundane tasks.”

Tasks that are costly, non-personal, and do not require creativity. This is where AI can make the biggest impact. Report creation, data entry, initial inquiry responses, meeting minutes, template emails — eliminate these “necessary tasks that don’t need to be done by humans” first.

3. Measure effectiveness within one month.

Assess how time and costs have changed before and after implementation. If there are no results, stop. If there are, expand the application. The key is to start small and make decisions based on numbers.

The 1/76 figure from Asana and the 96% cut from Sophos were numbers designed with intention from the outset. It was not a case of “we implemented AI and it happened to work well.” They decided on the costs to eliminate and targeted AI at those costs, resulting in those figures.

The strength of small and medium-sized enterprises lies in their speed of decision-making and proximity to the field. While large companies take six months to get approvals, they can start testing by next week. That speed itself will be the greatest competitive advantage in the AI era.

The world of model costs reduced to 1/76 is already here. The question is whether your company will ride that wave.