Conclusion First

When using AI to write code, expect to see monthly costs ranging from 200,000 to 500,000 yen “dissolve”.

Claude Code, Cursor, Devin, GitHub Copilot Workspace—AI coding agents are emerging one after another, creating an atmosphere where it seems that engineers may no longer be needed.

However, the voices coming from those who have actually used these tools tell a very different story.

“I was shocked when I saw the API bill.”

This is no exaggeration. When fully utilizing AI coding agents, the token consumption far exceeds human imagination. For small and medium-sized enterprises, this “invisible cost” can be fatal.

Why Token Consumption Explodes—A Structural Issue

First, let’s understand why the token consumption of AI coding agents can balloon to such levels.

Traditional tools like ChatGPT or Copilot offer “complementary” suggestions, proposing just a few lines of code. The token consumption per request is at most a few hundred to a few thousand. A monthly plan costing $20 is usually sufficient.

In contrast, agent-type tools like Claude Code and Devin are fundamentally different. They read files, comprehend entire codebases, make modifications across multiple files, run tests, and fix errors autonomously. Each time this loop is executed, tens of thousands to hundreds of thousands of tokens are consumed per task.

Let’s look at some specific numbers:

• Input token cost for Claude 3.5 Sonnet: approximately $3 per million tokens, output: about $15 per million tokens.
• For models like Claude 3.5 Opus class, the costs are several times higher.
• Tokens consumed by an agent for a single feature implementation: an average of 50,000 to 200,000 tokens (including context loading).
• Running 10 tasks a day results in: 500,000 to 2,000,000 tokens per day.
• Operating for 20 days a month leads to: 10 million to 40 million tokens per month.

When applying these figures to API costs, the range becomes 300,000 to 600,000 yen per month. It’s not uncommon for developers running 2 to 3 agents in parallel to exceed 1 million yen monthly.

One overseas indie developer posted on X that after fully utilizing Claude Code for a week, they incurred $1,000 (about 150,000 yen) in API costs. This translates to 600,000 yen monthly, easily surpassing the labor cost of one engineer in a regional SME (40,000 to 50,000 yen per month).

“Fast” Does Not Equal “Cheap”

There is a structural issue that many people overlook here.

The value provided by AI coding agents is “speed.” They can complete an implementation that takes a human three days in just 30 minutes. This is a fact and is indeed overwhelming.

However, being “fast” and being “cheap” are entirely different matters.

Human engineers are on a salary. Paying 400,000 yen a month means they work 160 hours a month, equating to 2,500 yen per hour. Regardless of speed, the cost remains fixed.

On the other hand, AI agents are billed based on usage. Even if they finish quickly, if they consume a large number of tokens, the costs can escalate. Moreover, agent-type tools consume tokens during the “thinking” process as well. The more trial and error involved, the higher the costs. For humans, “thinking for a moment” incurs no cost, but for AI, even thought processes come with a price.

In other words, the break-even point for AI agents is determined not by “speed” but by “token efficiency.”

Quality Issues Further Drive Up Costs

Adding to the cost problem is the issue of quality.

The code generated by AI may work, but there is a deep chasm between “working” and “production-ready.”

Here are some common issues encountered in the field:

1. Ignoring Edge Cases: It may pass normal cases but crashes on abnormal ones. Missing input validation, null references, race conditions—deploying without human review can lead to accidents.

1. Excessive Abstraction: In its attempt to write “clean code,” AI tends to introduce unnecessary design patterns. As a result, code is produced that no one in a small team can maintain.

1. Loss of Context: During long interactions, AI may forget initial instructions. The direction may shift midway. The more modifications are made, the more tokens are consumed, leading to a vicious cycle of declining quality.

1. Security Holes: AI carelessly embeds vulnerabilities that fall under OWASP’s Top 10. Missing SQL injection protections, hard-coded API keys—AI prioritizes “functionality” over “safety.”

When these rework costs are added up, the time that AI was supposed to have “saved” can easily vanish. Debugging code written by AI in 30 minutes can take a human three hours—this is no joke; it has happened in multiple environments.

Where is the Break-Even Point?

So, where is the line for small and medium-sized enterprises to “profit” from using AI coding agents?

Let’s do a simple calculation.

Assumptions:

• Engineer labor cost: 450,000 yen per month (for a mid-level engineer in a regional area)
• Monthly working hours: 160 hours
• Hourly rate: approximately 2,800 yen

AI Agent Costs:

• API costs: 150,000 to 500,000 yen per month (depending on usage frequency)
• Assuming 30-50% rework on the code generated by AI requires human review and correction.
• Review labor: 40 to 80 hours per month (25-50% of the engineer’s working hours)

Break-Even Calculation:

• Without AI: One engineer produces output for 160 hours a month → Cost: 450,000 yen
• With AI: API cost 300,000 yen + review labor (half a person) 225,000 yen = 525,000 yen
• Only if the output from AI increases by 1.5 times or more does it finally break even.

In other words, if the output with AI is less than 1.5 times, it actually results in increased costs.

In reality, while the “quantity” of AI output may increase, the amount of “usable output” often does not increase proportionately. If half of the generated code is discarded, the actual productivity improvement is only around 1.2 to 1.3 times. This results in a deficit.

What Should Small and Medium-Sized Enterprises Do?

This is not to say, “Don’t use AI.” It’s about how to use it.

1. Start with “Complementary” Rather Than “Agent” Type

In many cases, a monthly fee of $10 to $19 (about 1,500 to 3,000 yen) for Copilot is sufficient. Code completion, test code generation, documentation creation—these can all be adequately covered by a flat-rate complementary model. Agent-type tools that charge based on token consumption should be limited to tasks where ROI is clearly visible.

2. Visualize Token Consumption

Many development teams do not monitor API costs. This is akin to using a credit card without keeping a budget. Visualize token consumption and costs daily, and understand the unit cost per task. “Implementing this feature with AI cost 5,000 yen”—if you can’t manage at this level of granularity, cost optimization is impossible.

3. Categorize “What to Let AI Do”

There is a clear distinction between areas where AI excels and areas where it struggles.

Efficiently Handled by AI:

• Routine CRUD implementations
• Test code generation
• Refactoring suggestions
• Documentation and comments generation

Should Be Done by Humans:

• Architecture design
• Defining security requirements
• Core business logic
• Identifying edge cases

Failing to categorize this and “dumping everything on AI” will lead to token explosions and quality collapse.

4. Set a Monthly Cap

Simple but effective. Set a limit, such as “API costs for the month should not exceed 100,000 yen.” If it does, humans will write the code. This alone can prevent runaway costs. Claude Code’s API usage has a spend limit setting feature. It would be unwise not to use it.

The Real Question is Not “What Has Become Cheaper with AI”

After reading this, some may wonder, “Is AI ultimately expensive?”

However, that is not the essence of the matter.

What AI coding agents confront us with is the fact that “the value of writing code itself is decreasing.”

As the cost of writing code decreases, the market value of being able to write code also declines. Conversely, the value of the ability to decide “what to create” and the ability to “deliver what has been created” increases relatively.

For small and medium-sized enterprises, this is actually an opportunity. Large companies employ a vast number of engineers, and their labor costs weigh heavily as fixed expenses. Small and medium-sized enterprises can use AI as a “variable cost.” They can run it only when needed and stop when it’s not. This agility is something large corporations cannot replicate.

However, to leverage this agility, it is essential to accurately understand the cost structure and identify the break-even point.

“Using AI will make it cheaper” is an illusion. Only companies that can design how to use AI to profit will reap the benefits.

First, take a look at this month’s API bill. There, you will find hints about your company’s break-even point.