Why Corporate America is rethinking its AI spending

Corporate America is becoming more selective about how much it spends on artificial intelligence as companies discover that the most expensive models are not always necessary for routine work.

Businesses are increasingly combining premium systems from OpenAI and Anthropic with cheaper open-weight models, including products developed in China.

The shift is changing how companies measure AI spending, with cost becoming almost as important as raw model capability. Some businesses are also negotiating heavily subsidized usage from AI providers as competition intensifies.

Companies are moving away from one model for everything

Businesses that once prioritized access to the most powerful AI models are increasingly choosing different systems for different jobs. The approach allows companies to reserve expensive models for complex tasks while using cheaper alternatives for routine work.

Mike Saeks, a field chief technology officer at Cursor, said companies no longer need the most powerful models for relatively ordinary tasks. Cursor’s software allows customers to switch between models from multiple AI companies, making it easier for businesses to compare cost and performance.

The change represents a significant shift from the earlier focus on maximizing AI usage. Companies once viewed heavy token consumption as a sign that employees were making extensive use of AI. Now, businesses are paying closer attention to how many tokens they need and what they receive for that spending.

Lower-cost models are changing AI budgets

The economics of model selection become particularly important when AI systems are handling large workloads. Cursor recently tested the cost of rebuilding SQLite from scratch in Rust, using different combinations of AI models, and found major cost differences despite broadly similar outcomes.

Using GPT-5.5 as both planner and worker costs $10,565. Using Anthropic’s Opus 4.8 for planning and Cursor’s Composer 2.5 for implementation costs $1,339.

That gap illustrates why businesses are experimenting with model combinations instead of automatically selecting the most expensive system available. In Cursor’s experiment, the premium model handled planning and high-level decisions, while the less expensive model performed most of the implementation work.

Cursor field CTO Mike Saeks has also said that the best model for a task can now change multiple times a week, making flexibility increasingly important for companies managing AI costs.

Chinese AI models are becoming part of the equation

The push for cheaper AI has also created a geopolitical dimension. Many leading U.S. frontier models remain closed, while Chinese developers, including DeepSeek, Moonshot AI, MiniMax, and Z.AI, have become important sources of open-weight models that businesses can download, deploy, and customize.

Some U.S. companies remain cautious about Chinese models because of security, intellectual-property, and governance concerns.

Anthropic has accused DeepSeek, Moonshot, and MiniMax of using large-scale distillation campaigns to extract capabilities from Claude. OpenAI has separately accused DeepSeek of using distillation techniques to benefit from U.S. frontier models.

The U.S. policy debate remains unsettled. Trump administration officials have threatened targeted sanctions or other measures against Chinese companies if intellectual-property theft is established, while also expressing support for open-weight AI more broadly.

NVIDIA, Microsoft, Meta, Palantir, and other technology companies have backed an industry letter arguing against premature restrictions on open-weight models and warning that broad limits could weaken competition and innovation.

Meanwhile, U.S. companies are expanding their own open-weight offerings. Meta, NVIDIA, OpenAI, and other developers now provide open models, giving businesses alternatives to both Chinese open-weight systems and leading proprietary models.

AI companies are competing for customers

The growing focus on costs is also changing the relationship between AI providers and their corporate customers. Companies are receiving increasingly aggressive offers as OpenAI, Anthropic, and other providers compete to retain business.

Marty Kausas, chief executive of AI customer support company Pylon, said there is little customer loyalty as businesses compare models and pricing.

Pylon has received substantial amounts of free AI usage from vendors. Kausas estimated that the company had received about $1.6 million in free tokens from one vendor, $65,000 from another, and $10,000 from a third.

Such incentives show how AI providers are trying to establish long-term relationships with businesses, while companies have more leverage to negotiate pricing and usage arrangements.

The competition is particularly important as major AI companies prepare for potential public listings and continue to support valuations based on expectations of rapid growth.

Open models can reduce the cost of large AI operations

Zoom has used Meta’s Llama, an open-weight model, for 3 years and has reduced costs by fine-tuning the system, according to its chief technology officer, Xuedong Huang.

Zoom now uses a combination of Anthropic, OpenAI, and open models. Huang argues that combining several systems can yield a stronger overall result than relying on a single model for every task.

Hex, an AI data analytics company, has also seen increased adoption of open models among its customers. Chief executive Barry McCardel said that about 50% of Hex customers had adopted Kimi, a China-based model from Moonshot, into their workflows over the previous 2 weeks.

For companies that can customize models with their own data or optimize them for specific tasks, open-weight systems can offer another way to balance performance and cost.

Telnyx shifts workloads after an AI billing change

Telnyx provides another example of how quickly AI economics can change at scale. In the spring, the company was running about 1,000 AI agents using a top Anthropic model via a subscription that cost $200 per employee per month.

After Anthropic restricted the use of subscription credentials through third-party AI harnesses, Telnyx estimated that switching to pay-as-you-go access for the same usage could cost roughly $100,000 per day.

Telnyx subsequently shifted more work to open-weight models. The company reported running about 1,400 agents, using models from the Chinese developer Z.AI, at around $100 per agent per day.

Telnyx still uses premium models for specialized roles. Anthropic’s Fable model has been used to plan work, while open-weight models perform implementation, and OpenAI’s Sol model reviews the output.

The example shows how companies can distribute work among multiple providers rather than relying on one premium model for every stage of an AI workflow.

Companies are building hybrid AI strategies

The growing preference for model combinations is also appearing in legal technology. Harvey has post-trained GLM-5.2 for specialized legal workloads and has also experimented with systems that can call Anthropic’s Fable 5 when a task requires additional capability.

That approach reflects a broader shift in how businesses think about AI infrastructure. Instead of choosing a single provider and committing most workloads to it, companies can distribute tasks across several models based on performance, cost, and customization requirements.

The strategy can also reduce dependence on a single AI laboratory at a time when new models are rapidly reshaping the competitive landscape.

AI spending is shifting from scale to efficiency

The corporate AI market is not necessarily moving away from advanced models. Companies that need the highest level of intelligence can still justify paying premium prices. The change is that businesses are becoming more selective about where that premium is necessary.

For routine tasks, cheaper models can provide adequate performance. For complex reasoning, planning, or quality control, companies can reserve their most capable systems for the work that benefits most from them.

For Corporate America, the emerging strategy is not necessarily to abandon large AI investments but to allocate them more selectively. Businesses are increasingly matching models to individual workloads based on capability, customization, and cost.

TL;DR

• Companies are increasingly combining expensive frontier AI models with cheaper open-weight alternatives.
• Businesses are using premium models for complex tasks while less expensive systems handle routine implementation work.
• Cursor found that rebuilding SQLite cost $10,565 when GPT-5.5 handled both planning and implementation, compared with $1,339 when Anthropic’s Opus 4.8 handled planning, and Cursor’s Composer 2.5 handled implementation.
• Chinese open-weight models are gaining adoption despite ongoing security, intellectual-property, and geopolitical concerns.
• Pylon has received about $1.6 million in free AI usage from one vendor this year, according to its CEO.
• Zoom uses a combination of Anthropic, OpenAI, and open models after using and fine-tuning Meta’s Llama for 3 years.
• Hex CEO Barry McCardel said that about 50% of the company’s customers adopted Moonshot AI’s Kimi model into their workflows within 2 weeks.
• Telnyx shifted more workloads toward open-weight models after estimating that pay-as-you-go usage of Anthropic could cost roughly $100,000 per day.
• Companies are increasingly building hybrid AI systems that use different models according to the complexity, performance requirements, and cost of each task.

Don’t forget to follow us for more exclusive content.

If you liked this, you might also like:

• U.S. restrictions test China’s fast-growing humanoid robot industry
• Adult children rush to lock down parents’ savings as AI scams explode
• Quantum computing investors may want to watch these two rising stocks

This article was made with AI assistance and human editing.