Conclusion First

Forget about the price of APIs.

OpenAI has reduced the price of the GPT-5.6 API by 20%. Google Gemini Flash is now at one-tenth the cost of its predecessor. The valuation of the inference-specialized chip “Etched” has reached $21 billion, indicating that hardware is also aggressively driving down costs.

What this trend signifies is clear: AI inference costs are heading towards zero.

“Since it’s cheaper, let’s use AI”—those who think this way should take a step further. The real question is:

“When the cost of running AI becomes almost negligible, what will differentiate success?”

The answer is clear: data organization and business design. Whether small businesses can invest time and money in these areas will determine their future.

—

The Collapse of Inference Costs—What’s Happening in Numbers

First, let’s lay out the facts.

• GPT-4 (March 2023): Approximately $30 per million tokens input
• GPT-4o (May 2024): Approximately $5 (one-sixth)
• GPT-5.6 (2025): Further 20% price reduction
• Gemini 2.0 Flash: About one-tenth compared to the previous generation

In rough terms, inference costs have decreased to less than one-twentieth in just two years.

Consider a small business making 100,000 API calls a month. The inference costs that were around 300,000 yen per month in 2023 can now be reduced to under 15,000 yen. That’s a difference of nearly 3.5 million yen annually. This is no longer a “cost” but rather close to an “error”.

Moreover, competition in inference-specialized chips is intensifying. Etched boasts more than ten times the inference efficiency of existing GPUs, with a valuation of $21 billion. Groq is also achieving ultra-low latency inference with its proprietary chip, the LPU. They are breaking down costs at the semiconductor level.

Cash technologies have also evolved. Systems that cache frequently used prompt patterns to skip recalculation have been implemented, further lowering the effective API call costs. OpenAI’s Prompt Caching has reported up to 50% cost reduction during cache hits.

In short, the cost of “using AI itself” is no longer a variable in investment decisions.

So, what will become the variable?

—

What Becomes Valuable When Costs Approach Zero

This is the crux of the matter.

When the costs of technology dramatically decrease, the value of “something else” invariably increases. History has proven this.

• As printing costs approached zero → The value of content quality and editorial skills increased.
• As communication costs approached zero → The value of what to convey (information design) increased.
• As server costs approached zero → The value of what to host (service design) increased.

In a world where AI inference costs approach zero, the things that will gain value are “what to feed AI (data)” and “where to integrate AI (business design)”.

Conversely, if these two aspects remain weak and one merely says, “I’ve implemented AI for now,” it’s akin to trying to run a high-performance engine with water.

—

What Small Businesses Should Do “Right Now” #1: Data Organization

When you hear “data organization,” it might sound grandiose. However, what small businesses need is not the construction of a big data infrastructure.

What needs to be done is simple: “Organize the information scattered within the company into a format that AI can read.” That’s it.

To be specific:

• Write out the estimation logic in the heads of veteran employees into a spreadsheet.
• Organize three years’ worth of inquiry emails into folders and convert them into CSV.
• Scan paper manuals and run OCR to convert them into text.
• Standardize inconsistencies in product master listings (e.g., “㈱” vs. “株式会社”, mixed full-width and half-width).

Each of these tasks may seem mundane. However, whether or not you do them can change the quality of AI output drastically.

For instance, let’s take a local manufacturing company (30 employees) that we assisted. It took about 40 hours of work and approximately 200,000 yen in outsourcing costs to organize around 2,000 estimation data entries from Excel to CSV and correct inconsistencies.

As a result of connecting this data as a reference for RAG (Retrieval-Augmented Generation), the time required to create estimates was reduced from an average of 45 minutes per case to 8 minutes. When calculated over 60 estimates per month, that’s a savings of 37 hours per month, or about 450 hours annually. In terms of labor costs, that translates to over 1.5 million yen in annual benefits.

With an investment of 200,000 yen, the return is 1.5 million yen annually. The ROI is 750%.

The key point here is not the API costs. The monthly API costs are only a few thousand yen. What made the difference was whether or not they invested 200,000 yen in data organization.

—

What Small Businesses Should Do “Right Now” #2: Business Design

Alongside data organization, reviewing business design is equally important.

There’s a common failure pattern: “Keeping the current business flow as is and just adding AI.”

For example, consider this case:

> They implemented an AI chatbot for inquiry responses, but staff ended up checking every case. Instead of reducing labor, the checking workload increased.

This is a business design issue. If the AI output is designed with the assumption that “humans will check everything,” labor will not decrease; it may even increase.

The correct design is as follows:

1. For responses that AI can answer with over 90% confidence → Automatic replies (not seen by humans)
2. For those with 70-90% confidence → Draft generation + light human review before sending
3. For those with less than 70% confidence → Human handling

By simply dividing it into these three stages, the need for full checks is eliminated. In fact, a local e-commerce company (12 employees) that implemented this found that about 65% of inquiries fell into Category 1, resulting in customer support working hours reducing from 20 hours a week to 7 hours a week.

The cost of reviewing this business design was about 300,000 yen for two workshops and setup work. The monthly labor cost savings effect is about 250,000 yen. They recouped the investment in the second month.

Again, the API costs here are only a few thousand yen per month. The investment should be in the “design” aspect.

—

The Structure Where Small Businesses Can “Win” Because They Are Small

What happens when large corporations implement AI? It takes three months for approvals, two months for security reviews, two months for vendor selection. Six months to start the PoC. A year for full deployment. During that time, API costs may drop further.

Small businesses are different. If the CEO says, “Let’s do it,” they can start next week. In a company of 30 people, they can organize all employees’ data in a month. Changes to business flows can be shared in a morning meeting and implemented the next day.

The speed of decision-making and the closeness to the field create a structural advantage that large corporations cannot replicate.

In a world where AI inference costs approach zero, the ability to “use AI” will not be a differentiator. Anyone can use it. What will set businesses apart is “how quickly they can fit it into their operations.”

Here, small businesses have a significant advantage.

—

So, What Should Be Done?

Just three things:

1. Choose the most “person-dependent task” within your company by the end of this month.
Tasks that only veterans can do, customer response rules known only to specific individuals, or decision criteria that exist only in someone’s head. Write them down and digitize them. A budget of 100,000 to 300,000 yen is sufficient.

2. Design the “business flow” that incorporates AI into that task.
Clearly delineate between “what AI will do” and “what humans will do” rather than saying, “Leave everything to AI.” Spending 300,000 yen here can be recouped in two months.

3. Don’t waste time comparing API costs.
Whether GPT-5.6 is 20% cheaper or if Gemini is more affordable. Instead of spending an hour on such comparisons, use that hour to convert one business manual into text. If you have time to debate a few thousand yen difference, focus on improving your business by tens of thousands of yen.

—

The Collapse of Inference Costs is the “Equalization of the Starting Line”

The fact that AI inference costs are heading towards zero means that “using AI” will no longer provide a competitive advantage. Just as using electricity is not a differentiator when electricity becomes cheap.

What will set businesses apart is what they operate with electricity, or in this case, what they run with AI.

Determining that “what” is the data organization and business design.

I genuinely believe that this is a historic opportunity for local small businesses. They can use the same AI as large corporations at almost no cost. The only competition is the “speed at which they can integrate it into their operations.” And small businesses can do this overwhelmingly faster.

It’s not the time to be swayed by news of API price cuts. The company that starts organizing data today will win in six months.