After months of encouraging employees to use artificial intelligence as much as possible, major companies like Uber and Microsoft are now pumping the brakes, embracing a more strategic approach to AI spending that matches task complexity to model cost. The shift away from "tokenmaxxing" (using AI indiscriminately) toward deliberate model selection is reshaping how organizations think about their AI budgets in 2026.

What Is Model Routing and Why Does It Matter?

Model routing is the practice of directing different tasks to different AI models based on complexity and cost. Instead of defaulting to the latest, most expensive model for every request, companies are now asking: Does this task really need GPT-5.5, or would a cheaper alternative like Claude Fable or an open-source model work just as well ? Morgan Linton, chief technology officer of AI startup Bold Metrics, exemplifies this approach. He tells his 16 engineers which specific models to use for different tasks, allowing his team to "use the best stuff, but they're using it a lot more efficiently," he explained.

The economics are compelling. Coinbase CEO Brian Armstrong predicted in June that "80% of workloads will be running on 99% cheaper models within 12-18 months," with only the remaining 20% requiring cutting-edge models where "IQ maxxing is important". This forecast reflects a growing consensus that most routine tasks don't justify premium pricing.

Brian Armstrong

How Are Companies Implementing Smart Model Selection?

• Task-Based Routing: Companies assess whether a task requires advanced reasoning or can be handled by simpler, cheaper models. Tanvi Pisal, a user-experience designer at a major tech company, now designs interfaces in Figma first, then uses Claude only for functionality and flow, rather than brainstorming from scratch. This design-first approach "really helps me save tokens," she noted.
• Model Testing and Switching: Software engineer Alejandra Thomas runs tests on every new model to understand its strengths and weaknesses. "I try not to use the most expensive or advanced model just because it's available. For simple tasks, I always use lighter models or none at all," she stated.
• Automated Routing Platforms: Startups like OpenRouter, Rayline, and Fireworks are building software that automatically directs requests to the most cost-effective model. David Gilmore, who runs Rayline, noted that his tool "intercepts requests and determines whether they could go to cheaper, often open-source models." Usage of routing platforms has grown from around 1% of firms last year to 5% this year, according to Ramp's lead economist.
• Asking Models to Self-Assess: Some organizations are using a clever tactic: asking a cheaper model first whether a more expensive one would be needed for a task. "The models themselves are actually getting really good at assessing their own complexity," noted Spencer Yang, managing partner at investment firm BlockSpaceForce.

The shift reflects what behavioral economist Dan Ariely calls a "scarcity mindset." When companies impose token budgets, employees feel pressure to use them or lose them, similar to how people once tried to max out their monthly cell phone minutes. Token caps create "a psychology of waste if people don't reach their target," Ariely explained.

Why Did Companies Embrace Tokenmaxxing in the First Place?

The first half of 2026 was dominated by tokenmaxxing, a philosophy that encouraged maximum AI usage. The logic seemed sound: newer models reduce retries, supervision, and wasted effort. OpenAI's Kaylin Voss noted on LinkedIn that better models "reduce retries, supervision, and wasted effort". However, once companies reviewed their AI bills, the reality set in. The cost of using premium models for routine tasks became unsustainable, prompting a recalibration.

Chris Maconi, cofounder of AI startup Hechura, was never convinced by tokenmaxxing. He runs his company with a "human-in-the-loop" attitude and isn't setting up overnight bots to keep coding continuously. When he deployed an OpenClaw agent (a Mac Mini-based AI system known for burning through tokens with 24/7 use), he started with cheap Gemini models before switching to Anthropic's Haiku. "I'm not afraid to go and try some of these lower-end models to see if they can provide the intelligence that we need," Maconi said.

What Are the Barriers to Smarter Model Selection?

Despite the clear financial benefits, adoption of model routing remains uneven. Maconi attributes this partly to laziness. "People don't want to do the hard work of understanding which models are good at which things," he observed. "They just want to ride the hype train". This inertia means many organizations continue defaulting to expensive models even when cheaper alternatives would suffice.

The growing adoption of routing platforms suggests this barrier is eroding. As more companies experience sticker shock from their AI bills, the motivation to understand model trade-offs increases. The venture capital community has taken notice, with startups like OpenRouter receiving significant funding to automate these decisions.

How Does This Trend Affect AI Education and Ethics?

Beyond corporate cost-cutting, educational institutions are grappling with how to teach responsible AI use. Arkansas State University-Mountain Home published an updated guidebook in 2026 to help students and faculty navigate AI tools ethically. The guidebook addresses a real concern: nearly 60% of U.S. college students use AI tools like ChatGPT, Google Gemini, or Microsoft Copilot for classwork at least once a week, with roughly a fifth using it daily, according to a Gallup and Lumina Foundation poll.

Jessica Clanton, a science faculty member who led the guidebook's development, emphasized that "AI is a fact of life. It's here to stay, and like any tool, it can be used for good or evil". The latest revision incorporated feedback from over 80% of surveyed students who wanted clearer guidance on proper AI use. The guidebook emphasizes that AI should be a tool to support learning, not replace critical thinking.

Miranda Edwards, a nursing student and tutor at ASU-Mountain Home, noted that students shouldn't be "guessing" how much AI use is permitted. "A lot of students worry about plagiarism with AI," so the guide "eliminates a lot of frustration," she explained. The guidebook's approach mirrors the corporate shift toward intentional, purposeful AI use rather than blanket adoption.

The broader lesson from both corporate and educational settings is clear: AI's value lies not in ubiquitous use, but in strategic deployment. As models become cheaper and more specialized, the competitive advantage will go to organizations and individuals who understand which tool to use when, not those who simply use the most advanced option available.