LLM Prices Drop 97% in 18 Months—What Should SMEs Sell in the Era of ‘Zero AI Costs’?

Conclusion First: The Value of AI Itself Will Diminish. The cost of using LLMs at the GPT-3.5 level has plummeted by 97

By Kai

|

Related Articles

Conclusion First: The Value of AI Itself Will Diminish.

The cost of using LLMs at the GPT-3.5 level has plummeted by 97% in just 18 months.

When looking at the price per million tokens, it was about $20 at the beginning of 2023, and by late 2024, it is expected to be below $0.60. In simple terms, what used to cost 300,000 yen per month for AI processing is now less than 10,000 yen per month.

Those who see these numbers and think, “It’s cheaper now, so we should implement it too,” are half right and half off the mark.

The essence is this: Using AI itself is no longer a differentiator. Anyone can use the same model at nearly the same cost. Both large corporations and SMEs can hit the API and get the same GPT-4o response.

So, what will set them apart?

“How to redesign business operations based on AI”—that’s the only factor.

What Happens When Costs Drop by 97%

First, let’s clarify the structure.

When the cost of AI dramatically decreases, three things happen simultaneously.

1. The ‘Reasons Not to Do It’ Disappear

Previously, it was a reasonable decision to postpone implementation because it cost tens of thousands of yen per month. Now, you can start for just a few thousand yen per month. Even a company with five employees can use an AI chatbot for customer support. There’s no reason not to try it.

2. ‘We Implemented AI’ No Longer Differentiates

Competitors can do the same thing at the same cost. “Efficiency through AI” becomes a given, and it doesn’t generate a single yen in revenue by itself.

3. The Definition of Human Work Changes

This is the most significant change. If AI can work 24/7 for just 10,000 yen a month, what is the point of having humans do the same tasks? We need to fundamentally redraw the lines between “work that should be done by humans” and “work that should be delegated to AI.”

Here’s a question for you: When was the last time you reviewed your company’s workflow?

A World Where the Cost of Phone Support Drops from 3 Million Yen to 120,000 Yen Annually

Let’s look at a specific example.

Bland has announced its full-duplex voice model, “Tofu-1.” This model differs structurally from traditional voice AI.

Traditional systems operated on a multi-step pipeline: “voice → text conversion → LLM processing → text → voice synthesis.” Delays were introduced at each conversion step, making the conversation feel unnatural. People could tell over the phone, “Oh, this is AI.”

Tofu-1 processes voice directly. It thinks while listening and speaks while thinking, just like a human. Even in poor communication environments or when the other party’s speech is interrupted, it maintains context and responds.

Applying this to phone support for SMEs would look like this:

Item Traditional (2-part-time staff) AI Voice Agent
Monthly Cost About 250,000 yen (1,200 yen/hour × 8h × 2 staff × 21 days) About 10,000 yen (API costs + call charges)
Annual Cost About 3 million yen About 120,000 yen
Support Hours Weekdays 9 AM to 5 PM 24/7
Concurrent Handling Up to 2 cases Virtually unlimited
Quality Variation Human-dependent None

A cost reduction of 2.88 million yen annually. Moreover, the support hours expand more than threefold.

Of course, I’m not saying that every phone call should be replaced by AI. Human intervention is necessary for complaints and complex inquiries. However, consider how much of the overall volume consists of routine calls like “confirming business hours,” “changing reservations,” and “inquiries about fees.” In many companies, this should exceed 70%.

If you can handle that 70% for just 10,000 yen a month, you can focus all your efforts on the remaining 30%. This is the meaning of “business design based on AI.”

The Option of ‘Not Sending Data Outside’ with Self-Hosted AI

Another important trend for SMEs is the decreasing cost of self-hosted AI agents.

With services like “Smart Node,” you can build AI agents on your own servers or VPS. The monthly server costs are around 3,000 to 5,000 yen. This means you can avoid sending customer data or internal information to OpenAI’s servers.

This is quite significant for local SMEs.

The information that local companies handle—customer lists, contract terms with partners, region-specific business practices—often cannot be easily shared externally. I’ve heard many voices in the field saying, “Cloud AI is convenient, but sending our data outside is…”

With self-hosting, that psychological barrier lowers. You can have a company-specific AI assistant trained on your own information for just 5,000 yen a month. The ‘in-house AI’ that large corporations built for hundreds of thousands of yen is now within reach for SMEs.

A Structure Where SMEs Can ‘Win’ is Emerging

Now, let’s get to the main point.

The drop in AI costs is actually creating a structural change that favors SMEs.

What happens when large corporations implement AI? It takes three months for internal approvals, two months for vendor selection, two months for security reviews, and six months for company-wide deployment. It takes over a year. Moreover, customizations are outsourced, accumulating additional costs in the hundreds of thousands of yen.

For SMEs, if the president says, “Let’s do it,” they can move by next week. They can try a tool that costs 10,000 yen a month and stop it the following month if it doesn’t work. The speed of decision-making and the low cost of failure. This is a weapon unique to SMEs.

In the era when AI costs were high, only large corporations could use it. Now that costs are approaching zero, the competition has shifted to ‘who can establish AI-based workflows first.’ And in this game, small and agile organizations have a significant advantage.

So, What Should We Do?

I’ll say just three things.

1. Create a ‘List of Routine Tasks’ for Your Company by the End of This Month

Phone support, email replies, estimate creation, daily report compilation, invoice processing. First, write everything down. Then, mark the tasks that “do not require judgment.” That’s where you’ll start replacing with AI.

2. Start with Tools You Can Try for Under 10,000 Yen a Month

There’s no need to implement a large system all at once. Even ChatGPT Plus (at $20 a month) is fine. You can create an internal FAQ bot using no-code tools like Dify in less than a day. There are perspectives that can only be seen by ‘touching it first.’

3. Decide ‘What to Let AI Do’ Rather Than ‘What to Let Humans Do’

This is the most crucial question. After AI handles routine tasks, what will humans do with their freed-up time? Building relationships with customers, planning new businesses, improving on-site operations—that’s where the competitive edge for SMEs lies.

What to Sell in the Era of Zero AI Costs

Finally, to answer the question in the title.

In an era where AI costs approach zero, what SMEs should sell is not the “results of AI utilization.” That will become accessible to anyone.

What to sell is “human judgment and relationships that cannot be replaced by AI.”

Knowing local business practices. Matching faces with names of customers. Noticing, “That president seems a bit down lately.” Being able to rush over with just a phone call when needed.

These are things that no matter how cheap LLMs become, cannot be provided through an API.

Delegate everything that can be handled by AI to AI. Then, channel the freed-up time and costs into things that only humans can do.

That’s how SMEs can thrive in the ‘Zero AI Cost Era.’

The 97% drop in costs is not a threat. For SMEs, it’s the greatest opportunity yet. However, only those who take action can seize this opportunity.

POPULAR ARTICLES

Related Articles

POPULAR ARTICLES

JP JA US EN