As AI use rises but costs become burdensome, companies are actively trying to use AI as cost-effectively as possible.

More companies are trying to cut AI costs by mixing cheaper models as needed alongside high-performance models from OpenAI and Anthropic. In the process, router software that automatically selects cheaper models based on task complexity is drawing attention. Some companies also bring in open-source models and customise them to reduce costs.

The Information recently reported that Fin, which develops AI customer support agents, has been shifting from Anthropic’s Sonnet model to developing its own customised AI based on open-weight models since about a year ago. The company said the shift has picked up speed over the past few months.

Through this, Fin cut annual costs by hundreds of millions of dollars and significantly improved profit margins. Fin was recently sold to Salesforce for $3.6 billion.

Salesforce buys AI customer service startup Fin for $3.6 billion

Some companies, such as Meta and Uber, also impose limits on how much AI employees can use. The two companies once allowed employees unlimited AI use, but changed course and put in place restraints after exceeding their budgets.

Some also control which AI models employees can access. The idea is that they are separating employees who can use expensive models from those who cannot to cut costs.

Simply instructing AI models to think less can also reduce AI costs.

Enterprise software company UiPath uses prompt engineering to minimise the time AI models spend “thinking” before a task. It said it uses this approach especially for tasks it has done before. This led to cost reductions of more than 90 percent for some tasks.

Usage of high-performance AI models offered by Anthropic and OpenAI is still rising at an unprecedented level. If those models significantly outperform open-source models, it could help prevent users from leaving, but many assess that the gap between Anthropic and OpenAI’s advanced models and open-source models is narrowing at this point.

In particular, perceptions of Chinese open-source AI models have moved well beyond the view that they are merely usable at a cheap price. Some benchmarks have released results showing they outperform OpenAI or Anthropic.

How fast is Chinese AI catching up with the United States...opinions split in the tech world Global AI market sees a strong gust of Chinese AI armed with cost-effectiveness

OpenRouter, which provides a service that lets users choose among various AI models, said the share of tokens processed by open-source models rose to about 65 percent in the second week of June. That is a sharp increase from 34 percent in January and 26 percent in the same period last year. China’s DeepSeek and MiniMax models were particularly strong.

OpenRouter said small AI startups use many open-source models, but demand from large companies has also been increasing lately.

Factory Data, a software development tools company that launched its model router on June 1, is seeing a similar trend to OpenRouter. Factory said open-source model usage surged threefold from the previous month. Adobe, IT services company Ypro and payments company Adyen are using the Factory router.

The cost-saving effect of using open-source models instead of commercial models such as Anthropic is significant. Adam Sandman, CEO of software development company Inflectra, said switching from Anthropic’s model to Alibaba’s open-source model Qwen cut AI costs by 99 percent for product features that analyse images and videos in the software development process.

Coinbase, the largest cryptocurrency exchange in the United States, also cut AI spending by nearly half while increasing token usage by using Chinese open-source models.

According to Coinbase CEO Brian Armstrong (브라이언 암스트롱), the company deployed Z.ai GLM 5.2 and Moonshot Kimi 2.7 at the forefront.

Armstrong also shared that the company made it so engineers can clearly understand how many tokens they use. "Engineers can use as many tokens as they want on the model they want, but we made usage visible. The more you spend on AI, the more results you need to deliver," he said. He also stressed that "the goal is not the number of tokens, but reducing wasted tokens," adding, "the key is not to suppress usage but to build infrastructure that makes exponential growth sustainable."

Given that AI will be used more broadly in corporate environments, the strategic value of managing AI costs is expected to grow further.

Some say AI costs are different in nature from cloud and SaaS spending and require a different approach.

SiliconANGLE reported that Gartner Vice President and Analyst Marco Meinardi (마르코 메이나르디) said, "In the past, internal users and systems determined costs, but with AI, costs change depending on how customers use AI applications and what prompts they enter. How customers use AI apps affects a company's costs. It is a different problem, so it needs a different solution."