Paul Deraval, Cofounder & CEO of NinjaCat, is a software veteran with 20+ years driving innovation in martech, AI and agency growth.
We’re in the maxxing era. Looksmaxxing for those wanting to improve their appearance; proteinmaxxing for the diet-focused; moneymaxxing for the future rich; and bedmaxxing for the chronically exhausted. Then, of course, there is tokenmaxxing—the trend of maximal AI use to climb a corporate leaderboard or catch the AI wave at all costs.
While the productivity and product results didn’t materialize, the costs sure did. According to Fortune, “Uber burned through its entire 2026 AI budget in four months,” joining the nearly 70% of companies reporting AI cost overruns in 2026.
Unsurprisingly, executives and companies big and small are trying to rein in this spending. They are looking at which models cost less. It’s the right objective with the wrong approach.
The biggest AI costs rarely come from model selection alone. They come from operating AI without visibility. For example, according to TechSpot, Amazon integrated Anthropic’s Claude Sonnet model to match author information with product listings. The project ran for five months before the cost overrun was detected, ultimately generating a $1.8 million AI bill, 860% above its original budget. The incident highlighted how difficult it can be for enterprises to track and control AI spending when usage is metered by tokens.
The maxxing era is coming to a close. We are entering the value era, and the companies creating the most enduring value may not be using the best models, but they will know how to use their tools more intelligently. Here’s how you can make the shift.
1. Stop measuring tokens like they’re just another cloud bill.
Aware that AI won’t be free, most organizations are tracking their AI spend. They are measuring their cost per token, unsure what token usage actually represents but certain that high token usage equates to waste. It doesn’t.
Tokens reveal how AI is behaving. They indicate where models are spending the most effort, which workflows require excessive reasoning, where prompts lack clarity and where context forces unnecessary computation.
It’s like receiving a surprisingly high energy bill from running the air conditioner during the summer. It could be a sign that you need to use the air conditioner less, or that it has been running nonstop because a window was left open. Closing the window is probably a better solution than turning off the air conditioner.
Token usage is a diagnostic tool that can help your company use AI more effectively, not just limit overall expenditure.
2. Optimize the work before you optimize the model.
Nobody wants to be the next infamous headliner, so organizations have AI costs top of mind. It’s part of the reason open-weight models, from DeepSeek to Kimi K3, receive so much attention. Companies want to control costs, and these models promise a far cheaper alternative.
Leading with upfront cost is the wrong starting point. Instead, consider the work you are asking the AI to perform, understanding how much reasoning the work requires and the information the AI needs to succeed.
If cutting costs is the only goal, the cheapest model will probably achieve this outcome, but it may produce significantly worse results. Additionally, one study found that choosing cheaper models with higher error rates, even very modest ones, can easily erase the initial cost savings. Understand the work you need the AI to perform and optimize that task before selecting the model to run it.
3. Change how you think about models.
AI models are specialists, not generalists. Executives often think about models the same way they think about employee seniority. They assume that an expensive model must be used for important work, and a cheaper model is for lower-stakes tasks. In reality, different models are designed for different jobs.
Some excel at complex planning and pattern recognition, while others are built for scalable content generation or data classification and extraction. Running an advanced reasoning model on repetitive work wastes resources, while running a lightweight model on strategic analysis yields disappointing results.
Instead of running an advanced reasoning model on repetitive work and wasting resources, switch to a lightweight model purpose-built for high-volume execution. The most expensive model isn’t always the right choice, and matching the right model to the right task is the best way to maximize capability without incurring cost overruns.
Visibility is your AI competitive advantage.
Visibility is already a standard discipline across every other core business function. Advertising relies on campaign reporting to optimize conversion rates and return on spend, cloud infrastructure uses telemetry to monitor system health and uptime and finance uses data dashboards to track burn rates and prevent budget leaks. Too often, AI operations are opaque, even as businesses spend money to scale them enterprise-wide.
They don’t know which models are running, which workflows are creating value, which systems are struggling or which prompts require improvement.
Visibility changes this dynamic, allowing leaders to intentionally improve workflows rather than simply reacting to costs after the fact. The next competitive advantage in AI won’t come from having access to the newest model. The companies that outpace the competition will understand each required task before assigning a model, they will provide the context the AI needs to operate effectively and they will monitor how it performs, not just what it costs. They will do all this to refine workflows that improve performance and control costs without compromise.
Forbes Technology Council is an invitation-only community for world-class CIOs, CTOs and technology executives. Do I qualify?
