The AI Cost Collapse: Ultra-Cheap Tokens Reshape SaaS Margins

For two years, AI has been dominated by model benchmarks and one-upmanship.

Which model is smarter?

Which one writes better code?

Who had the best benchmark score?

Founders, investors, and business owners care about a different number: the cost of utilizing intelligence.

The simple question is: How much does it cost to utilize intelligence?

That cost is plummeting.

DeepSeek's V4 Flash API offers off-peak prices of $0.22 per million uncached input tokens and $0.66 per million output tokens. During off-peak hours cached input can drop to $0.007 cents per million tokens.

Google's Gemini 3.1 Pro costs $2 per million input tokens and $12 per million output tokens for prompts up to 200,000 tokens.

That is not only a technical detail.

This process transforms the economics of an AI company.

The AI cost collapse makes intelligence cheaper, presenting both opportunities and challenges for businesses building AI software.

The commodification of AI inference

One simple definition helps before we go further.

Inference occurs when an AI model is executed and an answer is generated.

Companies pay inference costs every time a customer interacts with an AI application, generates an image, or requests a task.

If we go generally, traditional software served another user easy costs remain extremely low.

Unlike purely reactive computer programs, just about any additional input you funnel into AI software takes more compute and tokens.

This creates a new kind of software business: the AI wrapper, which provides a product around a model.

A product around a model: an AI wrapper that wraps on top of an existing model; when you think in layers.

The company might provide:

a better interface;

specialized prompts;

templates;

workflow automation;

document handling;

integrations;

billing and user management.

This model can still work.

However, falling model prices are shifting the competitive landscape.

It's getting harder to ignore the price difference.

Consider a simple workload:

Consider a workload of 10 million input tokens and 2 million output tokens.

Using current off-peak prices for DeepSeek V4 Flash, the total cost is $3.52.

10M input × $0.22 = $2.20*

2M output × $0.66 = $1.32*

Total: $3.52

Now compare that with the price listed for a comparable context size in Gemini 3.1 Pro:

10M input × $2 = $20*

2M output × $12 = $24*

Total: $44

The actual economics depend on the workload, model behavior, caching context length, and provider pricing. However, the example highlights the magnitude of difference: with the same volume in tokens, you could be paying orders of magnitude more depending on the model selected for the API.

And pricing can change.

Companies should view these numbers as a snapshot for financial analysis, not definitive prices.

The crucial point is that a model is not always better.

A model isn't always better; AI model selection is a compute business consideration.

Why Are AI Costs Collapsing?

There are a number of forces pushing prices lower.

Smaller and more efficient models

Not every task requires the most powerful model.

A simple classification task may not need a frontier reasoning system.

Companies can run a small model for commonplace queries and reserve an expensive model for difficult cases.

Competition

Vying for developers and enterprise workloads, AI providers are going at it with a fury.

Price is a key factor, as multiple providers can perform similar tasks.

Open-weight models

Open-weight models offer an alternative, with weights released according to licenses.

Businesses can explore DeepSeek or Llama-family models, deciding when and where inference makes economic sense.

But there is a distinction:

However, open-weight doesn't mean free.

A company operating a model themselves essentially still incurs at least computing and GPUs, electricity or cloud infra, storage + networking + monitoring / engineering / maintenance costs to keep the thing running.

Why does cheap AI cause problems for SaaS firms?

That is where the story gets more exciting.

Why does cheap AI cause problems for SaaS firms?

And it is.

A $10 AI bill falling to $2, while customers pay $20, is a better margin.

But there is a second effect.

But there's a second effect: cheap AI makes it easier for others to compete.

Consider an AI writing company, for example — at $30/month.

An AI writing company at $30/month can be imitated by another firm using a less expensive model at a lower price point.

This happening with falling AI prices creates a paradox:

AI is cheaper to run for AI but essentially easier and copyable, even the basic products.

And that is the true SaaS margin dilemma.

But it is not just the case that inference costs are ruining margins.

The true SaaS margin dilemma is that underlying economic value from a shallow AI wrapper erodes as it becomes more accessible and commoditized.

The aforementioned new AI SaaS margin paradox

Traditional software-as-a-service (SaaS) companies build a lot of value around software.

AI applications are more and more placed on top of one other layer: the model.

That creates a stack:

AI model → API / infrastructure → AI application → end customer

If the model is cheaper and ubiquitous, the application layer needs other reasons for customers to pay.

That reason could be:

proprietary data;

specialized workflows;

industry knowledge;

integrations;

distribution;

automation;

customer relationships;

security;

compliance;

measurable business outcomes.

A general "AI assistant" is less difficult to exchange.

Conversely, an AI built into a company's workflow around finance, legal, logistics, health, or manufacturing is harder to replace.

This explains why the next generation of companies in AI should compete less on which model they use and more on the business problem they're trying to solve.

Pivoting toward open-weight models

There is also another strategic option in the form of open-weight models.

This might leave one of the below choices open to a company:

Proprietary API

The provider operates the infrastructure.

The business operates a pay per use model.

Benefits of managed DevOps – Use Less, Deploy Quickly and Lesser Management.

Hosted open-weight model

An openweight model runs on a third-party server, often with inference billed.

This can deliver some of the economic upside associated with open models without needing the company to own GPUs directly.

Self-hosted model

The model is run on the company's own infrastructure.

This allows for better control over the hardware but leads to operational overhead.

Self-hosting isn't without its costs, and the cheapest option isn't always to self-host.

The total cost of ownership, including engineering and infrastructure, must be worked out by any business.

A small startup with unpredictable traffic could be economically wiser to pay an API provider.

Dedicated or self-hosted inference should be seriously considered for a company with large predictable workloads.

Self-hosting never becomes cheaper past a certain threshold on volume of tokens, and there's no common figure that applies.

How Engineering Teams Can Reduce Costs of AI

There's no urgent need for companies to build their own GPU infrastructure yet.

Easier ways to cut AI costs exist.

1. Route Requests by Complexity

Route requests by complexity: use the most costly model only for the most complex tasks.

For example:

Simple task → low-cost model

Moderate task → mid-range model

Complex reasoning → premium model

The model used for a serious research task won't be the same as the one needed to summarize something in five lines upon customer request.

It isn't of course to sprinkle the cheapest model everywhere.

Don't use the cheapest model everywhere; use the lowest-cost model that can do the job accurately.

2. Cache repeated information to reduce costs.

Caching is one of the most obvious areas for current AI monetization.

A great example DeepSeek pricing illustrates this is; V4 Flash has an on-demand at $0.007 per million tokens off-peak for cached input (cached means just running the model so it has access to understanding input) -> where as uncached input will cost 0.22 cents you can see right how much of a diffence that would make in your aplication / processing,

One thing you can infer from this is that the caching here can provide significant gains for companies with a massive system prompt or long and repeated context.

But, naturally, how much can be cached depends entirely on the workload.

3. Reduce Unnecessary Tokens

Higher number of token normally corresponds to higher charge.

Companies should measure:

average input tokens;

average output tokens;

Cache repeated information to reduce costs.

retrieved documents;

unnecessary instructions;

failed requests;

unnecessary model calls.

The easiest AI cost reduction is sometimes:

Omnitemplate T2: Decaheadh WHAT NOT to SEND in the information Ill-harned by model→ DO NOT END somodo (that's why we decached)

4. Use Smaller Models for Routine Tasks

AI systems don't rely on a single model.

Multiple models are used in production systems:

- A cheap model for classification

- A cheap model for extraction

- A mid-tier model for summarization

- A premium model for complex reasoning

This technique is more popularly known as model routing.

This technique is known as model routing, which reformulates model selection as an ongoing optimization problem, balancing cost and quality.

Batch processing is suitable when speed isn't crucial.

Not all AI tasks require real-time responses.

processing thousands of documents;

generating reports overnight;

classifying customer records;

analyzing historical data;

creating bulk summaries.

Companies can leverage batch-processing functionality provided by model providers for these workloads.

The goal is simple:

The goal is simple: avoid paying for premium real-time economics for work that can wait.

6. Measure Cost Per Business Outcome

This is perhaps the most critical change.

Companies shouldn't only ask:

How much is it to buy one million tokens?

They should ask how much First Contact Resolution for a single customer task costs.

For example, one model might be less expensive per token but require a hundred times more retries.

Another model can be more costly but gets it right from the start, making it cheaper on a product basis.

This would make the second model cheaper on a product basis.

When thinking about AI economics, focus on the task and business-outcome level, not the token level.

The Future of AI SaaS May Not Be Having the Best Model on Paper

The deflation of intelligence is redefining what it means to gain a competitive edge.

A few years ago, the acesso to a powerful model itself could feel like such an advantage.

More and more, accessing a model is an infrastructure.

However, the moat more likely to be stronger is:

The stronger moat is data, workflow, distribution, integration, and trust.

Consider two companies.

A gorgeous AI chat bot — Company A

Company A has a beautiful AI chatbot, but Company B provides an AI system that integrates with internal documents, CRM, accounting software, and customer-support systems, automatically fulfilling business processes.

Company B, even with similar models, generates a more defensible value.

This is the trend of AI SaaS we're moving toward.

What This Means for Founders

For entrepreneurs, the lesson isn't "Don't build AI wrappers."

"Don't build AI wrappers.

It's: the model isn't the product; the wrapper is.

This is a good start—the wrapper.

It can help validate demand.

It validates demand and identifies customer problems faced by entrepreneurs.

However, a product will eventually need to build in something that is not easily replicable by competitors who use the same API.

That could be proprietary data.

It can be just a custom workflow.

It could be distribution.

A solution that serves an industry need.

Or perhaps a dramatically better user experience.

The best AI companies combine several advantages.

The Bottom Line

AI Cost Collapse: More Than A Pricing Narrative

It's shifting the economics of software.

The prices of current APIS show the difference between AI models clearly. Current API prices show the difference between AI models clearly: DeepSeek V4 Flash has the lowest off-peak token prices ever, while high-end models like Gemini 3.1 Pro can be several times the price per million tokens.

That creates two opposing dynamics for AI companies.

MARGIN, increased margins due to lower inference cost.

The Bottom Line

Basic AI products become easier to replicate — If inference costs are low, anyone can copy with minimal effort.

Why will SaaS margin compression in AI?

The winners aren't necessarily those using the cheapest model.

They're organizations that encapsulate low-cost intelligence with something else customers can't easily substitute.

The Bottom Line

The model is becoming cheaper.

The utility of the workflow has become more valuable.

Perhaps as AI intelligence continues to drop in price, the greatest business opportunity will be exploiting cheap AI to address expensive problems.

Exploiting cheap AI to address expensive problems

Frequently Asked Questions

The concept of ultra low cost AI models

Ultra low cost AI models are low-price, high-content-saturation AI systems (e.g. APIs) that consume vast amounts of text for cheaper prices compared to premium frontier models. Their low cost is especially significant for applications processing millions or billions of tokens.

Why will SaaS margin compression in AI?

When the economics of an AI application worsen, we say SaaS margins for that AI application are compressed because either model costs go up or competition increases pricing pressure or infrastructure costs. But shrinking prices for models also makes margins better, so the main problem is usually about disappearance of any differentiation among commodity AI products.

What is an AI wrapper?

AI wrapper refers to an application on top of a pre-trained AI model or web API. An AI wrapper is an application on top of a pre-trained AI model or web API that can add an interface, prompts, workflows, integrations, and other features without building the actual core model.

How much cheaper is DeepSeek than building proprietary AI models?

For some workloads, yes. The V4 Flash API from DeepSeek currently provides 0.22/million uncached input tokens and $0.66 million output token during twilight hours.

Is self-hosting an open-weight LLM the lowest cost option?

No: self-hosting cuts a few API fees but adds equivalent (or more) of GPU costs, infrastructure, and engineering plus close monitoring and operational costs. Organizations need to compare the TCO with commercial API pricing.

Will companies manage to cut their work with AI by 80%?

Companies that combine model routing, caching, smaller models, prompt optimization, and efficient infrastructure might end up with very large reductions. However, 80% is not a blanket promise. The actual savings depend on the architecture and the workload.

Now, what is the greatest chance for a company to tap into artificial intelligence?

With the drop in price of AI technology, businesses are now able to automate tasks that had previously been too expensive to do so. That may be the weakest opportunity, and it's likely that the greatest opportunities will arise from using cheaper intelligence applied to more expensive, repetitive, or specialist business problems.