The first half of 2026 was defined by a single directive across tech companies: use more AI. Employees were told to tokenmaxx, burning through API budgets on every task regardless of complexity. That era is over. Business Insider reports that companies from Uber to Microsoft are now imposing cost discipline, and the primary mechanism is model routing, not token caps.

The Shift from Volume to Routing

Morgan Linton, CTO of AI startup Bold Metrics, told Business Insider he now tells his 16 engineers exactly which models to use for which tasks: Claude Fable on low settings for one team, GPT-5.5 on high for another, Cursor with Composer 2.5 for a third. “My team is getting to use the best stuff, but they’re using it a lot more efficiently,” he said.

Coinbase CEO Brian Armstrong crystallized the math in a June 7 post on X, cited by Business Insider: “80% of workloads will be running on 99% cheaper models within 12-18 months.” The remaining 20% stays on frontier models where reasoning quality matters.

This is the modelmaxxing thesis: match model tier to task complexity. Route the hard problems up. Route everything else down.

Why This Hits Agent Builders Hardest

For someone writing a Slack message with ChatGPT, model routing is a nice-to-have. For anyone running autonomous agents 24/7, it determines whether the operation is financially viable.

Chris Maconi, cofounder of Huntsville-based AI startup Hechura, told Business Insider he never bought the tokenmaxxing premise. When he set up his OpenClaw agent, 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,” he said.

The economics are straightforward. An always-on agent making hundreds of API calls per day will burn through a flat token budget in hours if pointed at a frontier model. The same agent, routing routine file reads and status checks to Haiku while escalating complex reasoning to Opus or GPT-5.5, can run indefinitely at a fraction of the cost.

Uber COO Andrew Macdonald said publicly that justifying AI spending is getting harder. Pylon’s CEO declared the tokenmaxxing era “coming to an end.” When the CFOs start asking questions, blanket API budgets are the first thing to get cut.

The Scarcity Psychology

Dan Ariely, behavioral economics professor at Duke University, told Business Insider that token budgets create a “model of scarcity where people can’t use as much as they want.” He compared it to the early days of mobile phone plans with limited minutes: people would make calls at the end of the month just to avoid wasting their allocation, then switch providers to avoid overage charges.

The parallel for agent builders is direct. A hard token cap incentivizes gaming the budget rather than optimizing the work. Model routing removes the cap entirely and replaces it with a routing decision: does this task need Opus, or will Haiku do?

The Infrastructure Layer Forming Underneath

The article mentions that tools are emerging to automate these routing decisions, removing the need for engineers to manually assign models to tasks. This is the infrastructure play forming in real time: model routers that sit between the agent and the provider, analyzing task complexity and cost-per-quality tradeoffs automatically.

For multi-agent deployments, where a single orchestrator spawns dozens of sub-agents for parallel work, the routing layer becomes the primary cost control surface. The question is no longer “how many tokens did we use?” but “did each task use the cheapest model that could complete it successfully?”

The Pricing Pressure Loop

This shift pressures model providers from both directions. Frontier models lose volume as routine tasks get downgraded. Cheap models face a race to the bottom as routing makes switching trivial. The winners are providers who can offer a spectrum of models at different price points within a single API, reducing friction for builders who want to route dynamically without managing multiple provider integrations.

For agent operators specifically, the implication is that cost optimization is now an architectural decision, not a budgeting one. The agent that routes intelligently will outperform the agent with a bigger token budget pointed at a single model. Efficiency beats scale.