Why Businesses Are Choosing Cheaper AI Models Over the Best Ones

The Enterprise AI Race Is Starting to Look Different

The biggest AI model is not necessarily the model businesses want to buy. New spending data from Ramp suggests that companies are increasingly choosing AI models based on price, practical performance, and value rather than automatically moving to the newest flagship model. 

The clearest example is Anthropic's Fable 5. Despite being positioned as the company's most capable model, Fable 5 accounted for only about 6% of tokens purchased from Anthropic and 11.4% of spending on Anthropic models in its first full month tracked by Ramp. 

That result is important because it challenges a basic assumption of the AI market: that customers will naturally pay more for the strongest available model.

Fable 5 Is Powerful, But Businesses Have a Budget

Ramp's August AI Index describes Fable 5 as the most performant model in its comparison, but also the most expensive, at roughly $10 per million tokens. OpenAI's GPT-5.6 Sol was priced at about half that level in Ramp's comparison. 

The spending data shows the difference clearly. GPT-5.6 Sol represented around 25% of OpenAI's token usage and 23% of its spending in the same period. Ramp calculated that Fable 5 generated approximately 75% as much model-attributed spending as GPT-5.6 Sol. 

That does not prove GPT-5.6 Sol is a better model overall. It shows something different: business customers may consider a cheaper model good enough for a much larger share of their workloads.

Why "Good Enough" Is Becoming a Powerful AI Strategy

For a company processing millions of AI requests, the difference between an excellent model and a slightly less capable model can become expensive very quickly.

A business might use its most capable model for difficult research, complex reasoning, or specialized development tasks. But using that same model for every email draft, classification job, document summary, customer-support response, or routine coding request may provide little additional value.

This creates a natural model hierarchy. Companies can reserve expensive frontier models for the tasks that actually need them and use cheaper alternatives everywhere else.

The Numbers Reveal a Bigger Enterprise Trend

Ramp's broader data covers more than 70,000 U.S. businesses using its payment and expense-management platform. In July, 43.5% of those businesses were paying for Anthropic subscriptions or tokens, compared with 39.7% for OpenAI. 

Anthropic therefore remains extremely important in enterprise AI. The interesting development is that its most expensive model is not automatically receiving the majority of customer spending.

Ramp also found that the overall share of its business customers paying for AI reached nearly 56% in July, up from just over 50% in March. That suggests the enterprise AI market itself continues to expand even while customers become more selective about individual models. 

Cheaper Models Are Changing the Competition

The shift is not limited to OpenAI and Anthropic. Ramp reported that model-serving platforms providing access to open-source and Chinese-developed models reached 6.1% of AI-adopting businesses in July, continuing a gradual increase.

This creates a more complicated market than a simple two-company race.

Businesses can increasingly choose between premium proprietary models, cheaper models from the same provider, competing AI labs, open models, and platforms that route workloads across several models.

The result is that AI companies are competing on economics as much as intelligence.

The Most Capable Model May Not Win Every Task

AI benchmarks can make model competition look straightforward. A leaderboard places one model above another, and it is tempting to assume that the highest-ranked system should also receive the most commercial usage.

Real businesses do not work that way.

A model can be technically superior while still being a poor choice for a particular workload if its additional capability costs significantly more than the productivity improvement it provides.

For example, if a cheaper model completes a routine classification task with nearly the same accuracy, there may be little financial justification for using the premium model. The difference becomes meaningful only when the additional intelligence changes the outcome enough to justify the extra cost.

AI Spending Is Becoming a Routing Problem

This could push companies toward multi-model AI architectures. Instead of choosing one AI provider for everything, organizations can route different requests to different models.

A simple task might go to a low-cost model. A difficult reasoning task could be sent to a premium model. Sensitive workloads could use an approved private deployment. Coding tasks might use a model optimized for software engineering.

In that environment, the winning AI platform may not be the company with the single strongest model. It could be the platform that helps businesses select the right model automatically while controlling cost, latency, security, and quality.

Why This Matters for AI Companies

The spending data creates a difficult challenge for AI labs. Training increasingly capable models requires enormous investments in computing infrastructure, research talent, data, and energy. Those costs are easier to justify when customers are willing to pay premium prices for premium intelligence.

But if customers increasingly choose cheaper models, AI companies may have to find ways to lower inference costs or demonstrate much larger productivity gains before enterprises accept higher prices.

Ramp's own conclusion is particularly revealing: Fable 5 may represent a new upper boundary for what businesses are willing to spend on an AI model without a sufficiently large performance advantage. 

OpenAI Is Also Feeling the Same Pressure

The trend should not be interpreted as an Anthropic-only problem. OpenAI is also competing in a market where businesses can switch models quickly.

Ramp's data showed Anthropic ahead among its tracked businesses in July, but TechCrunch reported that OpenAI was growing faster among that group during the third quarter to date. The same data showed that businesses are willing to move between providers as new models appear. 

That volatility means neither company can assume that enterprise customers will remain loyal simply because they have the strongest model at a particular moment.

The Hidden Metric Is Cost Per Completed Task

There is an important limitation to interpreting model spending data. Paying less for a model does not automatically mean a business is saving money.

A cheaper model that requires multiple retries, extensive human correction, or additional software to produce an acceptable result can ultimately cost more than a premium model that completes the task correctly on the first attempt.

The metric businesses ultimately need is therefore not simply cost per million tokens. It is closer to cost per successfully completed task.

That measurement is harder to calculate, but it provides a much more useful way to compare AI systems in production.

What Companies Should Consider When Choosing an AI Model

Businesses evaluating AI models should look beyond benchmark rankings and advertised intelligence. The better questions are practical.

• How accurately does the model complete the company's real workloads?
• How much does each successful task cost?
• How frequently does the model require retries or human correction?
• Does the model meet security and data-retention requirements?
• How quickly does it respond?
• Can workloads be routed between different models?
• Does the model provide enough additional value to justify its premium price?

This approach can prevent companies from spending heavily on capabilities their employees rarely need.

What the Shift Means for Developers

Developers are likely to encounter this model-selection problem directly. A team might use a high-end reasoning model for architecture and debugging but a cheaper model for boilerplate code, documentation, test generation, or simple transformations.

AI development tools could increasingly handle this routing automatically. Instead of asking developers to choose a model manually for every request, an AI platform could estimate the difficulty of a task and select an appropriate model based on cost and expected quality.

That would turn model selection into an infrastructure problem rather than a user-interface decision.

Frequently Asked Questions

Why are businesses not automatically choosing the most powerful AI model?

The most capable model can also be the most expensive. If a cheaper model performs sufficiently well for a task, businesses may prefer the lower-cost option.

How much enterprise spending went to Fable 5?

Ramp reported that Fable 5 represented 11.4% of spending on Anthropic models in its first full month tracked, while accounting for 6% of identified Anthropic token purchases. 

Does this mean Fable 5 is a bad AI model?

No. Low spending share does not establish that a model is technically poor. It indicates that businesses may not consider its additional capability worth the higher price for many workloads.

Are companies still adopting AI?

Yes. Ramp reported that nearly 56% of its tracked businesses were paying for AI in July, showing continued expansion even as customers become more selective about models. 

What is likely to happen next?

The market is likely to move toward a mixture of premium models, cheaper specialized models, open models, and automated routing systems that select models according to task requirements.

The AI Market May Be Entering Its Value Era

The latest enterprise spending data suggests that the AI market is entering a more mature phase. During the early model race, the central question was often which company could build the most powerful system.

Now businesses are asking a harder question: Is the extra intelligence worth the extra money?

That change could have major consequences. AI labs may still compete aggressively to build increasingly capable models, but commercial success will depend on whether those capabilities translate into measurable value for customers.

The next generation of AI winners may therefore not be determined by benchmark scores alone. They may be determined by who can deliver the best combination of intelligence, reliability, speed, security, and cost.