The Era of Zero-Cost Models Has Arrived

To get straight to the point: the price of AI models is heading towards zero. More accurately, it is already nearly zero.

Alibaba’s open-source model “Qwen” has surpassed 3 billion downloads, overtaking Meta’s Llama to become the most downloaded open-source LLM in the world. The latest Qwen3 offers eight sizes ranging from 0.6B to 235B, and the 27B model can run on a notebook PC with 16GB of memory. The cost is zero, and the license is Apache 2.0, allowing for free commercial use.

On the other hand, it has been reported that payment giant Stripe is negotiating to acquire AI gateway company OpenRouter for about $7 billion (approximately 1 trillion yen). OpenRouter provides a routing service that allows users to switch between multiple AI models via a single API. The infrastructure for “delivering and utilizing models” has been valued at 1 trillion yen, rather than the models themselves.

When we juxtapose these two pieces of news, a structure becomes apparent.

The value of models is decreasing, while the value of the infrastructure for “delivering and utilizing” models is increasing.

Is this good news for SMEs, or is it a new pitfall?

A Shift in Costs—From 3 Million Yen to 50,000 Yen and Beyond

Let’s recall the time just two years ago, right after the launch of ChatGPT. The API usage fee for GPT-4 was $0.03 per 1K tokens for input and $0.06 for output. If you used 1 million tokens a month, the API alone would cost tens of thousands of yen. Customization and fine-tuning often required initial costs of 3 to 5 million yen.

What about now?

• Qwen3-27B: Free. Runs locally. Extremely cheap via API.
• Google Gemini 2.5 Flash: Free tier available. Even paid options are a fraction of GPT-4.
• DeepSeek-R1: Open-source. High performance specialized for inference.
• Llama 4: Provided by Meta. Commercial use allowed.

A glance at OpenRouter’s pricing makes it clear. Models with performance comparable to GPT-4 can be used for $0.1 to $0.5 per 1 million tokens. This is less than a tenth of the cost two years ago. If run locally, it can literally cost zero.

In other words, using AI itself used to be a cost. Now it’s different. Not using AI has become the cost.

So, what should SMEs be spending their money on?

The Cost of “Choosing” Becomes the New Bottleneck

With models becoming free, a troublesome issue has emerged: there are too many choices.

Hugging Face has over a million models available. Even just the models connected by OpenRouter number in the hundreds. New models are released almost weekly, and benchmark rankings are constantly changing.

Large companies have dedicated ML engineers who can compare benchmarks, validate with their own data, and select the optimal model. SMEs lack those resources. In a company with five employees, asking someone to compare Qwen3, Llama4, and Gemini is simply unrealistic.

This leads to two patterns:

Pattern A: Unable to choose and thus stopping. Overwhelmed by information, they end up implementing nothing. This is the worst-case scenario.

Pattern B: Just using the paid plan of ChatGPT for now. This isn’t bad, but it means continuously paying 20,000 yen per month multiplied by the number of users. For ten people, that’s 200,000 yen a month, or 2.4 million yen a year. Many tasks could be replaced with free models.

Both are a waste.

Three Things SMEs Should Really Invest In

In a world where model costs are approaching zero, what should SMEs be paying for? Specifically, there are three things.

1. Invest in the Ability to “Decompose Tasks”

AI models are tools. Even if the cost of tools is zero, it’s meaningless if you don’t know “what to create.”

What SMEs should first do is inventory their operations and identify “tasks that can be replaced by AI.” This isn’t a technical issue; it’s a business issue.

Here are some examples:

• Creating estimates: Automatically generate drafts using Qwen3-4B based on past estimate data. Time required per item reduces from 30 minutes to 3 minutes.
• Responding to inquiries: Vectorize FAQ data and use RAG (Retrieval-Augmented Generation) for automatic responses. Of 100 inquiries a month, 70 can be handled automatically.
• Summarizing and analyzing daily reports: Summarize daily reports written by employees with LLM and extract trends. The manager’s review time reduces from 1 hour a day to 10 minutes.

All of these can be done sufficiently with free models. What’s needed is the discerning ability to judge, “This task can be delegated to AI.”

2. Invest in “Systematization”

Personalization is the greatest enemy of SMEs. As long as there are tasks that “only Tanaka can do,” that company will not scale.

The true value of AI lies in breaking down personalization.

Transform the know-how in Tanaka’s mind into prompts. Feed that into an LLM so that anyone can produce outputs of the same quality. This is systematization.

The specific steps are simple:

1. Interview veteran employees to document their decision-making criteria.
2. Incorporate that documentation into system prompts.
3. New employees use those prompts to carry out tasks.
4. Provide feedback on results to improve the prompts.

You don’t need expensive AI tools to run this cycle. If you set up a model locally using Ollama, the running cost is just the electricity for the PC. A world where you spend a few hundred yen a month.

What you should invest in is the “human time” to design and maintain this system. If you outsource, an AI utilization consultant for 50,000 to 100,000 yen a month is sufficient. That’s 600,000 to 1.2 million yen a year, which can be cheaper than full company adoption of ChatGPT.

3. Invest in “Organizing Company Data”

This is the most overlooked point.

Models can be obtained for free. However, to unleash a model’s performance in your business, you need your own data. Past estimates, customer interaction histories, product manuals, internal rules—without these organized, even the highest-performing models can only provide “generic answers.”

Conversely, your company’s data becomes the “moat” for SMEs. No matter how high-performing a model a large company has, it doesn’t possess your company’s 20 years of customer interaction know-how.

The costs associated with organizing data are actually not that high.

• Converting past documents to PDF/text: Can be handled during employee downtime.
• Building a vector database: Free options like ChromaDB or FAISS are available.
• Constructing an RAG pipeline: Using LangChain or LlamaIndex, you can create something functional in a few days.

Even if you outsource, it would only cost about 500,000 to 1 million yen. If done in-house, it’s mostly just labor costs.

The Acquisition of OpenRouter Indicates the “Next Battlefield”

Consider why Stripe is willing to pay $7 billion for OpenRouter.

OpenRouter does not create models; it is a service that “connects” them. It routes user requests to the optimal model. For Stripe, this is significant for securing payment infrastructure in the AI era.

What can be inferred from this is that the competition to “create models” is winding down, while the competition for “how to deliver and utilize those models” is beginning.

The message for SMEs is clear: the era of worrying about performance differences between models is over. Regardless of which model is used, decent performance will be achieved. The differentiating factor will be “what to do with that model” and “how to incorporate it into operations.”

The Question of “Is This Enough?”

Finally, I want to ask candidly.

Is your company still leaving AI as something “not well understood”?

Is one person using a 20,000 yen ChatGPT subscription and saying, “We’re doing AI too”?

In the world, free models downloaded 3 billion times are running on notebook PCs. API prices are dropping every month. The $7 billion acquisition indicates that this trend is irreversible.

The price of models will reach zero. However, the cost of “the ability to utilize them” will not.

What SMEs should do right now is only three things:

1. Choose one of your company’s operations and try replacing it with a free model. Just one. It can be an estimate or a daily report.
2. Document that process so that anyone can replicate it. Avoid personalization.
3. Start organizing your company’s data. This will become your competitive advantage in six months or a year.

You don’t need expensive tools or complicated technologies. What’s needed is just the judgment to “start moving first.”

In a world where the price of models has reached zero, the most costly thing is “doing nothing.”