Conclusion First: The Price of AI is Starting to Break

In the same week that Anthropic’s annual revenue reached $65 billion, a project that runs a 0.4B parameter LLM on 7 ESP32 boards (costing around $30 total) gained attention, and OpenAI reduced the API price for GPT-5.6 by 20%.

When these three pieces of news are lined up, a single structure emerges.

The era of paying for “AI intelligence” is coming to an end.

While the revenue from high-performance models is growing, this is primarily because a few large companies are using them extensively, not because they are chosen for being “expensive is good.” Rather, on the ground, the sentiment is rapidly spreading that “as long as it works and is cheap, that’s enough.”

Is this good news for small and medium enterprises, or the beginning of confusion?

The answer depends on whether they have a “map.”

—

The Fact That a LLM Runs on 7 ESP32s for About $30

First, let’s clarify the story about the ESP32.

The ESP32 is a microcontroller board that can be purchased for about $4 to $5 each. It comes with both Wi-Fi and Bluetooth capabilities. Typically used for IoT sensors and simple controls, a project has emerged that clusters 7 of these to perform inference with a 0.4B parameter LLM.

To be honest, what can be done with 0.4B parameters is limited. It is less than one-hundredth the scale of GPT-4 class. Complex text generation and advanced reasoning are not possible.

However, it is premature to dismiss it as “useless.”

What matters is the fact that a LLM can run on $30 hardware.

A year ago, if you wanted to run an LLM yourself, you needed at least an NVIDIA GPU. An A100 would cost over a million yen. Now it’s $30. Of course, the performance is not comparable. But there is a difference as vast as between 0 and 1 when it comes to whether it can “run” or “not run.”

Consider this in the context of small and medium enterprises:

• Classifying simple anomaly patterns on a factory inspection line
• Automated responses to standard inquiries at a store reception
• Natural language input for picking instructions in a warehouse

For tasks that involve “limited contexts and limited responses,” 0.4B parameters could be practically sufficient. Moreover, it requires no internet connection, incurs no cloud API costs, and has a monthly cost of zero. Only the hardware cost of $30.

This is not competing with “expensive AI.” It is fundamentally breaking the very concept of ‘the cost of using AI.’

—

The 20% Price Cut on GPT-5.6 – Its True Meaning

OpenAI has reduced the API price for GPT-5.6 by 20%. This might be easily brushed off as “Oh, another price cut,” but structurally, it has a different meaning.

Looking back at OpenAI’s price trends over the past year:

• When GPT-4 Turbo was introduced: Price cut to about one-third compared to GPT-4
• When GPT-4o was introduced: Further 50% price cut
• This time for GPT-5.6: Another 20% price cut

Cumulatively, the cost of APIs with equivalent or better performance has dropped by about 80% in a year and a half.

For small and medium enterprises using 100,000 tokens per month, what used to cost around 30,000 yen a month a year and a half ago is now about 6,000 yen a month. That’s a difference of about 290,000 yen annually. This amount is close to the monthly salary of one part-time worker in regional small and medium enterprises.

Moreover, the reason for the price cut is “competition.” Google, Meta (Llama), and Chinese companies are all lowering their prices. This trend is unlikely to stop.

The clear conclusion that can be drawn from this is that “AI API costs” are no longer a major factor in investment decisions.

The time spent hesitating over a few thousand yen in API costs is far more costly.

—

The $65 Billion of Anthropic Shows “Another Reality”

On the other hand, there is news that Anthropic’s annual revenue has reached $65 billion (about 9.7 trillion yen). This is a 38% increase from $47 billion in May, in just two months.

While saying “cheap AI wins,” the sales of high-performance models are skyrocketing. It may seem contradictory, but it is not.

The ones paying are large corporations. Companies with enterprise contracts are driving Anthropic’s revenue by using APIs at a monthly cost of tens of millions to hundreds of millions of yen.

In other words, the AI market is becoming polarized.

• Upper Layer: Large companies exhaustively use top-performance models at a monthly cost of tens of millions of yen
• Lower Layer: Small and medium enterprises and individuals use “sufficient performance” cheaply for a monthly cost of a few thousand to tens of thousands of yen

The middle ground is disappearing. AI services that are “somewhat expensive and somewhat intelligent” will find themselves squeezed from both above and below.

What small and medium enterprises should glean from this is one thing.

“Do not try to implement the same AI as large companies in the same way they do.”

—

Three Things Small and Medium Enterprises Should Do Right Now

Let’s move beyond abstract discussions and get specific about what to do.

1. Create Your Own “AI Cost Map” in 30 Minutes

List all the AI tools and services you are currently using (or considering). Write down how much you are paying monthly, what you are using them for, and what problems would arise if they were gone.

Just creating this list can reveal cases where “what you are doing for 30,000 yen a month can be replaced by a 500 yen API.” Conversely, it will also clarify areas where “only high-performance models are necessary.”

Discussing “AI implementation” without a map will only lead to getting lost.

2. Start One Experiment That Can Be Tried for Under 5,000 Yen

The API cost for GPT-5.6 is just a few thousand yen even if you use 100,000 tokens per month. Claude Sonnet is similar. Experiment with one thing you can do with this.

For example:

• Automatically generating drafts for estimates (semi-automating 20 estimates a month)
• Summarizing daily reports and automatically converting them into weekly reports
• Classifying customer inquiry emails and suggesting response templates

All of these can be started for under 5,000 yen a month. If the results are not satisfactory, you can stop. The “experience” gained from a 5,000 yen experiment is worth more than a 500,000 yen consulting report.

3. Consider “Edge AI”

The ESP32 example is extreme, but the direction is correct. “Edge AI,” which runs AI on local devices without sending data to the cloud, offers three benefits for small and medium enterprises.

• Zero Monthly Costs: No API costs incurred
• Data Remains Internal: Reduced security risks
• No Internet Required: Usable in environments with poor communication, such as factories and warehouses

I’m not saying to buy an ESP32 right away. However, by installing Ollama on a Raspberry Pi 5 (about 10,000 yen), you can run a 7B parameter model locally. For zero monthly costs, you can have a decent Japanese chatbot in-house. This is sufficiently valuable as an “experiment.”

—

Don’t Become a Seller of “Expensive AI”; Win as a User

Finally, I want to convey the most important point.

The essence of what happened this week is that “the commoditization of AI is happening faster than expected.”

Performance that was “cutting-edge” a year ago can now be obtained for just a few thousand yen. It will become even cheaper in six months. In this flow, the stance that small and medium enterprises should take is clear.

Do not invest in AI itself. Invest in “what will change in your company with AI.”

Model performance becomes obsolete in six months. However, the “business flow” that reduced the time to create estimates from 30 minutes to 5 minutes using AI will remain, even if the model changes. The “process” that systematized the inspection know-how that was previously dependent on individuals will remain even if the responsible person leaves.

Spend money and time on what will remain. Engage with what will disappear at minimal cost.

Now that the price of AI is breaking down, it is the perfect timing for small and medium enterprises to move faster than large companies by leveraging “affordability.”

If a vendor comes with a proposal saying, “Implementing AI will cost 3 million yen,” in an era where a LLM runs on a $30 microcontroller, ask them this:

“How much of that 3 million yen will still hold value a year from now?”

If they cannot answer, you can discard that proposal.