To borrow from the British idiom, companies no longer want to use an AI sledgehammer to crack a nut.

It seems like just yesterday that ChatGPT came on the scene, changing everything from how people proofread their resumes and make grocery lists to the way they plan vacations. As more tools like Claude and Gemini entered the picture, so did so-called “tokenmaxxing” in which employees competed, often at the behest of managers, to use the most tokens on the job. (Tokens are the units used to measure AI use and determine payment.)

But pulling out Anthropic and OpenAI’s offerings for basic tasks can be akin to calling an Uber just to go around the block or an electrician to change your light bulb. The task can be done much more simply, and for a lot less money, another way. Enter “modelmaxxing,” which entails routing tasks so that they’re done by the most efficient and cost-effective model.

“The AI race is no longer just about building the best model,” said Soumen Mandal, principal analyst at Counterpoint Research. “While performance will remain important for both enterprises and consumers, the right balance of capability, cost and deployment flexibility will ultimately determine the winners in the AI market.”

Those winners increasingly include Chinese AI companies that are offering capable models for a fraction of the cost of their US rivals. Just this month, Chinese AI firm Moonshot launched Kimi K3, a model that outperformed all rivals except Anthropic’s Claude Fable 5 and OpenAI’s GPT-5.6 just before Alibaba previewed its Qwen3.8 Max model, which it claimed is even more powerful.

Model Money Management

In June, Coinbase CEO Brian Armstrong took to X to explain how to keep AI spend flat while token usage grows, including by better routing. “In our custom harnesses, we preprocess prompts and route to the best model for the job, considering cache hits and model pricing,” Armstrong wrote.

It’s no surprise that cost is top of mind for companies now that there are options for their AI. “This is the first year where AI spend has been a big topic,” OpenAI CEO Sam Altman told CNBC at the annual Allen & Co. Sun Valley Conference. “And all of a sudden, it’s a very big topic. Everyone’s asking what we can do to help reduce spend or increase value.”

As companies consider their options, Chinese large language models such as DeepSeek, Doubao, GLM, Hunyuan, Kimi and Qwen have made major strides in the US. An IDC survey shows that 47% of 260 US decision-makers at companies with more than 1,000 employees reported using made-in-China models for at least one use case, and 20% reported extensive use. In a separate survey from late last year of roughly 100 respondents, 73% reported using or testing AI-driven routing, and 72% reported using or testing automated routing based on input or metadata analysis.

While the firm can’t conclude from the data that the demand for model selection and routing is primarily driven by cost savings, the pricing advantage of Chinese AI models is substantial. Chinese models like DeepSeek and Qwen typically charge less than $5 per million output tokens while US frontier models like GPT 5.5 and Claude Opus are priced about $25 to $30, Mandal said.

System Rerouting

Model routing is becoming a more important part of the AI stack, said Adam Swick, head of strategy at OpenRouter, a marketplace for AI models. A common pattern is to use a highly capable model to plan a complex task, break it into steps and determine what needs to happen, then use lower-cost models to execute the more routine parts of that plan.

“We have seen meaningful growth in open-weight models, particularly for coding, agentic workflows, reasoning and high-volume applications where cost efficiency matters,” Swick said. (Open-weight models, like those released by DeepSeek, keep their core components public, while closed-weight ones like those from Anthropic and OpenAI don’t.) “Developers are increasingly evaluating models based on how well they perform for a specific workload rather than relying solely on the lab or brand behind them.”

Closed model options tend to be more accurate, but more expensive and slower. Open-weight ones can be less accurate, but they’re also fast, cheap and offer developers more control. Plus, with a specialized open-weight model, an enterprise can get a product that is fast, cheap and accurate for a specific task, said Iz Beltagy, research scientist at the Allen Institute for AI (Ai2).

“It’s true that this model won’t be good at everything else, but who cares if it is good at the one task they care about?” Beltagy said.

Risky Business

But companies have more to think about than cost. There’s also a security concern. The US recently added Chinese e-commerce company Alibaba, internet search provider Baidu and other tech giants to a list of companies it says are aiding the Chinese military. The US Defense Department won’t be allowed to contract with the firms.

“Being public about using a Chinese LLM in the US market triggers, at best, suspicion; at worst, reputational damage and procurement pushback,” Igor Marchal, a vice president analyst at Gartner wrote in a recent report. “Operationally, if you do not lock down how a model runs, you risk leaking sensitive corporate data or siphoning intellectual property into untrusted infrastructure.”

Swick said it’s important to distinguish between the model lab that develops a model and the provider that actually hosts and serves it from a data center. Data handling, security, compliance, geography and reliability can depend heavily on the provider and deployment architecture, not simply on where the underlying model was originally developed, he added.

Many developers pull “fine-tuned derivatives” from open-source repositories, Marchal wrote, not knowing that the underlying architecture was built on a Chinese model base and “introducing compliance and shadow AI potential headaches.”

The Performance Gap

The US is still ahead when it comes to figuring out where technology is heading, pushing the frontier and answering the really difficult questions, Beltagy said. (All the major capabilities, such as coding and agents, were introduced by US frontier labs and reproduced by everyone else.) The US also has a unique strength in fully open AI, not only in releasing models, but also in the training data, code, checkpoints, evaluations and methodology, so that anyone can understand, reproduce, and build on the work, he added. That allows the continued building of a vibrant ecosystem that’s not locked behind closed doors.

But Chinese models have made significant improvements in recent years in coding, multilingual capabilities, mathematics and productivity tasks, helping narrow the performance gap with leading US frontier models, Mandal said.

“Improvements in model efficiency without sacrificing quality, the availability of open-weight models and government support for sovereign AI development are helping Chinese AI companies achieve economies of scale,” he added.

What’s Next?

The fact that tokenmaxxing made headlines just months ago and is already being replaced by the next trend shows just how fast things move in the world of AI models.

In the future, Mandal said AI models will be commoditized for standard workloads, and long-term differentiation will come from how AI is deployed rather than from the models themselves. The value proposition will shift toward applications, enterprise software and infrastructure.

That means the next phase of competition will be defined by commercial viability (as in the balance of performance, cost, capability and deployment flexibility) rather than benchmark performance.

“Sovereign AI investments will reshape the competitive landscape, while the multi-model approach will create regional leaders rather than a single global winner,” he added. “As AI adoption is still in its early stages, we expect value creation to extend beyond foundation models to cloud providers, semiconductor companies, enterprise software vendors and AI application developers.”