The Era Where One Engineer’s Work Can Be Done for 50,000 Yen a Month

First, let’s look at the numbers.

Cognition Inc. has announced its coding model “SWE-2,” which boasts a 64% reduction in cost compared to the previous model (Fable 5.1). In benchmarks (FrontierCode 1.1 Main), it achieved a score of 50%.

What does this mean?

The “cost of having AI write code” has dropped to a level that is no longer comparable to hiring a human engineer.

Specifically, How Much Has It Become?

It may be a bit rough, but let’s consider the numbers with a practical perspective.

Previously, fully utilizing an AI coding agent could cost between 200,000 to 300,000 yen per month just for API usage fees. Applying the 64% cost reduction of SWE-2 directly, this translates to 70,000 to 100,000 yen per month. That’s an annual cost of 840,000 to 1,200,000 yen.

On the other hand, hiring one engineer in a Japanese small or medium enterprise costs around 6,000,000 to 8,000,000 yen annually, including salary, social insurance, and recruitment costs. Even in rural areas, it doesn’t drop below 4,000,000 yen.

1,000,000 yen per year vs. 4,000,000 to 8,000,000 yen per year.

Looking at this difference, can one confidently say, “Humans are still better”? Of course, AI can’t do everything. However, when it comes to routine coding tasks—such as test creation, bug fixing, and refactoring existing code—AI is already faster and cheaper.

What Happens When You “Delegate the Work of Three People to AI”?

Imagine a software development company with about ten employees.

Seven engineers, two salespeople, and one administrative staff. Personnel costs alone amount to around 50 million yen annually. For a regional contract development company, sales might be around 60 to 70 million yen, with profits being tight.

Now, let’s say we replace the routine work of three of the seven engineers with AI. This doesn’t mean firing those three. Instead, AI absorbs the “workload” of those three, allowing them to focus on higher-level design and client interactions.

AI cost: 100,000 yen/month × 3 lines = 300,000 yen/month, or 3,600,000 yen annually.
Cost of work for three humans: equivalent to 18,000,000 yen annually.

Difference: 14,400,000 yen.

This 14,400,000 yen directly impacts profits. Alternatively, it can be invested in new business ventures. Growth can occur without hiring. For small and medium enterprises in rural areas, having an extra 14,400,000 yen can be the dividing line between survival and failure.

However, Just Being “Cheaper” Isn’t Enough

Let’s pour some cold water on this.

Even if costs decrease, who will guarantee the quality of the code produced by AI? Unless this issue is resolved, small and medium enterprises will be too scared to use it.

Large companies have dedicated QA teams and systems for code reviews. However, small and medium enterprises often lack the resources to validate AI-generated code. It’s a similar structure to “outsourcing that is cheap but whose quality is unknown.”

There’s a noteworthy movement here: the QA platform called “ClientCoded.”

This tool automatically generates synthetic testing environments for the results produced by AI agents and evaluates performance in real-time. In other words, it’s a system where “one AI inspects the work of another AI.”

Until now, quality management for AI was based on the premise of “human visual checks.” This is gradually shifting to SLO (Service Level Objective) based automated management.

What changes?

The need to have people dedicated to using AI will disappear.

This may seem subtle, but it’s a decisive change. The main reason small and medium enterprises couldn’t adopt AI was that they lacked personnel who could effectively utilize it. If quality management is automated, that barrier will drop significantly.

A New Management Metric: “AI Labor Costs”

Recently, an interesting trend has emerged in AI spending data from U.S. companies. AI spending per employee has decreased by about 10%.

At first glance, this seems contradictory. While the adoption of AI is accelerating, spending per person is decreasing.

The reason is simple. The speed at which the cost of AI itself is decreasing is outpacing the speed of its adoption.

In other words, even though more business operations are utilizing AI, the total cost is decreasing. This is reminiscent of the dawn of cloud computing. It was cheaper to use AWS than to maintain your own servers—this structural change is now happening in the realm of “labor costs.”

Small and medium enterprise owners should start to pay attention to the item “AI labor costs.” They should compare human labor costs with AI running costs and consider optimal allocations for each business operation. This will become a fundamental operation in management after 2025.

The Reversal Structure That Allows Small and Medium Enterprises to Compete with Large Corporations

Now, let’s get to the main point.

For large corporations to adopt AI, they must go through security reviews, internal approvals, proof of concepts (PoC), and full-scale deployment… which takes at least six months to a year. There are inter-departmental adjustments, integration with existing systems, and discussions with labor unions. The process is slow.

What about small and medium enterprises? If the president says, “Let’s do it,” they can start moving next week. In a company of ten, it only takes 30 minutes to explain to everyone.

If the cost of AI is 100,000 yen per month, the risk of “trying it out” is almost zero.

Moreover, the decline in AI costs will continue. The 64% reduction of SWE-2 is likely to drop even further in six months. The longer one waits, the cheaper it becomes, but while waiting, competitors may adopt it first.

This is where small and medium enterprises can find their competitive edge.

• Quick decision-making
• Low cost of experimentation
• Small impact from failures
• Ability to deploy company-wide all at once

While large corporations spend a year on PoC, small and medium enterprises can achieve results in actual operations. This speed difference is the greatest weapon small and medium enterprises have in the age of AI.

So, What Should Be Done?

I’ll say just three things.

1. First, decide on one task to delegate to AI.
Don’t try to do everything at once. It could be automatic test generation, summarizing meeting minutes, or processing invoices—anything is fine. Just pick one and start it next week.

2. Redefine “the work that should be done by humans.”
If AI can write code, the value of engineers will no longer be “writing code.” It will be about understanding customer challenges, translating them into designs, and giving correct instructions to AI. The value of humans lies in these upstream processes. Hiring criteria will also change. It will be an era of hiring “people who can achieve results using AI” rather than “people who can write code.”

3. Monitor costs every month.
Track AI API costs, reduced man-hours, and quality scores. Follow these numbers every month. Make decisions based on data, not intuition. With SLO-based quality management tools now available, this is entirely feasible.

This Trend Will Not Stop

The cost of coding AI has dropped by 64%. Automation of quality management has begun. AI spending per employee continues to decrease.

These are not separate pieces of news. “The cost of hiring AI is falling below the cost of hiring humans”—everything is moving simultaneously toward that critical point.

A company of ten in a rural area can produce the same output as a company of one hundred in Tokyo. Such reversals are no longer mere fantasies.

I want to pose a question. What tasks in your company can be delegated to AI starting next week? Whether you have an answer to that question could determine your performance a year from now.