Imagine if every trip could be made in a Formula 1 car. You still probably wouldn’t drive one to buy groceries.
Yes, it’s faster and more powerful. But the job does not require that much performance. The same principle applies to AI. Not every task needs the most powerful model, just as more AI usage does not automatically create more value.
Previously, in AI Doesn’t Read Words Like You Do: F1 Car, Tokens and AI Usage, we explored how tokens make AI usage measurable. But once companies could count AI usage, a new question emerged: should more usage actually be the goal?
From Tokenmaxxing to Valuemaxxing
Earlier in 2026, tokenmaxxing became a buzzword across companies. Companies were pushing employees to use more AI, with some even tracking token consumption as a sign of adoption. Accenture reportedly went as far as warning employees they could risk missing promotions if they failed to demonstrate sufficient AI adoption.
Then the bills started adding up. Microsoft, after rolling out Claude Code to thousands of employees, began cancelling most licences after token costs reportedly ran through its annual AI budget months ahead of schedule.
That is the flaw in tokenmaxxing: it can reward activity rather than outcomes. Basic tasks such as converting PDFs into slides, repeated attempts to fix a poor answer, or a task that takes longer to check than to do manually can all consume more tokens without improving the result.
The better question is what we might call valuemaxxing: flipping the question from “How much AI are we using?” to “What did that usage actually achieve?”
Not Every Task Needs an F1 Car
Using the most powerful AI model for every task is like paying for performance the journey does not require.
A smaller, cheaper model may be enough to classify documents, extract basic information or answer routine questions, whereas a more capable model may justify its higher cost for complex analysis or tasks where getting the answer right on the first attempt matters.
The opposite can also be true. Choosing the cheapest model can backfire if employees have to retry prompts, correct mistakes or spend more time reviewing the output. What matters is the full cost of getting a usable result, not simply the price of each token.
Model choice is only one part of the equation. Managing costs can also mean reducing how much information the model processes, or deciding whether the task needs a written answer rather than a simple label or decision.
Other Ways to Manage Token Use
One approach is model routing. Microsoft Foundry, for example, can direct each request to a suitable model based on factors such as complexity, quality and cost, helping businesses avoid paying for more AI capability than the task actually requires.
Another approach is to minimise processing the same background information repeatedly. Memory layers like Mem0 allow AI models to retain relevant information from earlier interactions, so less context needs to be sent again with every request.
More recently, a different type of AI model has emerged that could offer another way to manage these costs. JEV, developed by TypeSafe AI, still processes input tokens, but instead of generating sentences token by token, it turns unstructured information into a fast, structured decision. For example, if an employee reports that they cannot access the company VPN, JEV can simply tag the request as “IT Support: Account Access.”
That changes the token equation. General-purpose models like ChatGPT or Claude are often used to generate a full response, while JEV can stop at the decision itself. For businesses handling large volumes of routine background tasks, that could mean fewer output tokens and less processing for jobs where a written response is not actually needed.
Spend Smarter, Not More
Tokens have given businesses something tangible to count. But once AI usage becomes measurable, it is easy to mistake the metric for the goal.
More tokens may mean employees are experimenting, or that they are retrying weak outputs, reprocessing the same context and using expensive models for ordinary work. The goal is not to use less AI for the sake of it, but to know which tasks deserve the F1 car, and which are just a grocery run.
A better measure is the outcome: whether AI saves time, produces better work or helps people make better decisions.
The tools will evolve, but the principle will remain the same. That is the shift from tokenmaxxing to valuemaxxing. Tokens may be the currency of AI usage, but value comes from understanding what each token is buying and whether the outcome is worth the cost.
Acknowledgements:
Thank you to the Sunway iLabs team for their invaluable contribution and insights in preparing this article.
References
Amnee, M. AI doesn’t read words like you do: F1 car, tokens and AI usage. FutureX Insights.
ashokkarania. Beyond tokens: Rethinking AI economics with Microsoft foundry. Microsoft. Microsoft Partner Community.
Bellan, R. The token Bill comes due: Inside the industry scramble to manage AI’s runaway costs | TechCrunch.
Griffiths, B. D. BCG official says companies ‘need to start the pump’ on AI tokens. Business Insider.
Keary, T. Why ‘tokenmaxxing’ is out and ‘valuemaxxing’ is in. Forbes.
Ropek, L. Companies are scrambling to stop employees from maxing out AI budgets with small tasks. TechCrunch.
Schmelzer, R. Why everyone is talking about Jev, the AI that doesn’t chat. Forbes.