AI Costs Are Starting to Break Down

Let’s get straight to the point.

The API price for GPT-5.6 has dropped by 20%. In the same week, the open-source LLM “Ox Alpha” achieved benchmark scores comparable to GPT-5.6.

The significance of these two events happening simultaneously is profound. The rationale for “paying a premium for commercial models” is rapidly diminishing. In other words, this is the week when vendor lock-in in AI is structurally beginning to collapse.

For small businesses, this is not a distant concern; it is a matter of decision-making this month.

—

What Happened: A Look at the Numbers

Price Cut for GPT-5.6

OpenAI has reduced the API usage fee for GPT-5.6 by 20%. Specifically, the cost has dropped from approximately $30 to about $24 per 1,000 credits.

For a company using 100,000 credits a month, the monthly cost would decrease from $3,000 (about 450,000 yen) to $2,400 (about 360,000 yen), resulting in an annual savings of about 1,080,000 yen.

For large corporations, this might be a rounding error. However, for a small business with 20 employees that was paying 450,000 yen a month for API costs, an annual saving of 1,080,000 yen is equivalent to one employee’s bonus. This is not a trivial amount.

The Emergence of Ox Alpha

Another significant development is that the open-source LLM “Ox Alpha” has recorded benchmark scores equivalent to GPT-5.6.

Being open-source means that the API usage fee is zero. If run on in-house servers or cloud GPU instances, the only costs incurred are infrastructure costs. Preliminary calculations show that running the same processing volume on AWS GPU instances can result in monthly costs that are about one-third to one-fifth of those for commercial APIs.

In other words, the monthly cost of 450,000 yen could shrink to between 90,000 and 150,000 yen.

—

The Essence Is Not Just a Price Cut, But the Emergence of Choice

What we should consider here is not just that “prices have gone down.”

It is a structural change that has eliminated the reasons to be tied to specific vendors.

Until now, AI adoption has effectively been a choice between OpenAI, Anthropic, or Google. Once prompts were developed, workflows established, and systems integrated into the organization, the switching costs skyrocketed. This is vendor lock-in.

However, now that open-source has reached performance levels comparable to commercial models, “models are becoming interchangeable components.”

What does this mean?

• Negotiation power is created. Simply being able to say, “I will switch to Ox Alpha” changes the dynamics of negotiations with commercial vendors.
• Risk of failure can be distributed. If a specific API goes down, it is now feasible to switch to a different model.
• Fine-tuning with proprietary data is possible. With open-source, fine-tuning to match industry-specific terminology and workflows can be done freely. Commercial APIs often come with many restrictions.

The third point is particularly significant for small businesses. Large corporations can cover a sufficient range of operations with generic models, but for small businesses competing in niche industries, specialized models trained on proprietary data become a powerful asset.

—

A Case Study: Disassembling a 223-Node Agent Graph

There is a concrete example. A company replaced a 223-node agent graph (a complex workflow involving multiple AI agents) built on a commercial API with a simpler configuration based on an open-source LLM.

The results were as follows:

• Number of nodes: Reduced from 223 to 12
• Estimated monthly API costs: Reduced by 70-80%
• Time to identify causes of failures: Reduced from several hours to a few minutes
• Staff required for maintenance: Reduced from 2 dedicated personnel to 1 part-time person

Having 223 nodes means “billing can occur at 223 points, and failures can happen at 223 points.” Reducing this to 12 was made possible because model performance improved, allowing for more processing in a single inference.

The same results can be achieved without a complex system. This is both a cost reduction and a way to eliminate dependency on specific individuals. Only the person who built the 223-node system could understand it, but with 12 nodes, handover becomes feasible.

For small businesses, the greatest risk is “if the person in charge leaves, it’s over.” If a simpler configuration can be adopted, that alone adds value.

—

So, What Should Small Businesses Do This Month?

There’s no need for abstract suggestions like “let’s rethink our DX strategy.” Let’s get specific about what to do.

Step 1: Visualize Current AI Costs (Time Required: 1 Hour)

First, open this month’s API usage invoice. Understand how much you are paying monthly and how much each process costs. Surprisingly, many companies do not do this.

Numbers to check:

• Total monthly API costs
• Cost per processing unit (how much it costs to handle one inquiry)
• Top 3 most costly processes

Step 2: Determine “Switchability” (Time Required: Half a Day)

Answer the following three questions:

1. Is the prompt dependent on model-specific features? → If it’s a generic prompt, it will likely work almost as is with a different model.
2. Is the output format dependent on a specific API format? → If it’s a standard format like JSON, switching will be easy.
3. Does the monthly cost exceed 100,000 yen? → If it does, the cost-benefit analysis for switching becomes relevant.

If the answer to all three questions is “yes,” it’s worth starting a pilot test within this month.

Step 3: Test on a Small Scale (Time Required: 1-2 Days)

A full switch is not necessary. Try running just the one most costly process using Ox Alpha or another open-source model.

Compare on just three points:

• Output quality: Is it acceptable compared to the commercial model?
• Processing speed: Does it disrupt business operations?
• Cost: How much does it decrease?

If the quality is sufficient for business needs, there’s no need to pay a high price for a model that scores 100. Being able to determine that “this task doesn’t require a 100-point model” is the first step toward cost reduction.

Step 4: Redesign with a “Multi-Model Assumption”

If the pilot yields positive results, change your design philosophy moving forward. Don’t bet everything on a single model. Choose the optimal model for each process and maintain a switchable structure.

This is not technically difficult. By simply parameterizing the API endpoints, switching models becomes significantly easier.

—

What Will Happen Next

The recent price cuts and the rise of open-source are just the beginning.

In six months, even more powerful open-source models will emerge. Commercial models will either have to lower their prices further or differentiate themselves through added value (security, SLA, specialization).

The cost of AI models will approach zero.

At that point, competitive advantage will shift from “which model you are using” to “how much proprietary data and business processes you can feed into the model.”

Models are components. Components become cheaper. When components become cheaper, the differentiating factor will be how you use those components.

Small businesses may not have the same volume of data as large corporations. However, they possess “rich data” about specific industries and customers. Photos from construction sites held by a local construction company, processing condition know-how from a small factory, property information held by a regional real estate company — these are data that large corporations cannot easily obtain.

In a world where AI models are cheaper, the most valuable asset will be “the unique data that feeds into those models.”

What needs to be done this month is clear. First, understand your AI costs. Next, test just one process using open-source. Then, based on the assumption that “models are interchangeable components,” start organizing your true assets — data and business know-how.