After going through the first half of the year of “as much as possible use of AI”, enterprises started to focus more on cost control. Several entrepreneurs, engineers and designers indicated that they were allocating different models according to the complexity of the task, handing over complex tasks to stronger and more expensive front-line models, and giving repetitive or simple tasks to cheaper models.
This approach is replacing the previously popular “tokenmaxing”. The latter emphasizes the use of as many models as possible to increase team frequency and output. But when businesses saw the AI bill rising, they started to move to more detailed cost management.
From multiple models to fine distribution
AI Founding Company Bold Metrics Chief Technical Officer Morgan Linton stated that he would regularly inform the engineering team about the models to be used in different missions. Some teams use Claude Fable 's low-end configuration, while some tasks are assigned to GPT-5.5 high-end configurations. And there's a team in Cursor that works with Composer 2.5 and says the results are stable enough.
In his view, a clear model division of labour would eliminate the need for teams to rely on strict token ceilings and would also improve overall efficiency.
Chris Maconi, co-founder of Hechura, also stated that he did not agree with the unrestricted extension of the model. In the case of OpenClaw, which is an AI agent that can operate autonomously over a long period of time, such tools tend to consume large amounts of token. When he is actually deployed, he will try a cheaper Gemini model and switch it to the Haiku of Anthropic.
The design and development team started the screening first.
Some non-engineered positions are also adjusting their use. Tanvi Pisal, a large plant user experience designer, said that she would initially make brainstorms and product demand files directly from Claude, starting from scratch, which consumed a lot of token, but not the best move.
She then completed the interface design in Figma and handed over the screenshot to Claude so that the model could complete the functionality and process while retaining the interface. She'll start with the company's ChatGPT business version, and then give more mature content to Claude to produce a more complete document.
Alejandra Thomas, a software engineer and technologist in New York, said she would test the direction of every new model. For simple tasks, she usually gives priority to lighter models, even without them.
The practice of AI Scoot CEO Ed Stevens is closer to a phased assessment: the team selects a model for a few months before deciding whether to continue using it or to move to an updated but more cost-effective alternative.
Model router tool starting to heat up
As enterprises become more cost-oriented, the tools dedicated to “model roads” begin to receive attention. Such software will determine the complexity of the task and then distribute requests to different models, which may also include open source models.
According to Rayline’s director, David Gilmore, many clients initially preferred high-priced programmes for fear of missing the latest model, but often retrenched their budgets when they saw the API bill.
Ramp Chief Economist Ara Kharazian has indicated that the proportion of businesses using model route platforms is increasing. His statistics from last year show that about 1 per cent of businesses use these tools, and this year this percentage has risen to 5 per cent.
The San Francisco Investment Agency BlockSpaceForce currently uses OpenRouter, Fireworks and Together AI. According to its management partner Spencer Yang, users can even allow cheap models to judge whether the task really requires more expensive model processing, as the model itself is becoming increasingly good at identifying the complexity of the task.
From business practice, AI's use focus is shifting from “as much as possible” to “use the model in the right place”. In the context of budget tightening, model switching and routeing capabilities are emerging as new approaches to team cost control.