Around the end of 2025, beginning of 2026, right when Claude Code had its first real boom moment, people started throwing around the term token maxxing. If you are in tech, or even tech-adjacent, you know exactly what this means. Token maxxing is basically the idea that you should use as many tokens as possible to squeeze the maximum output from AI models. Long-running tasks. Computer use. Multiple teams of agents. Squads running in parallel. Basically: turn your work into a token-burning machine and hope productivity comes out the other side.
You can say whatever you want — maybe you can even argue that this logic is correct, and that we really do need to burn a stupid amount of tokens to get the best out of these LLMs — but I think this was one of the greatest psyops ever pulled by AI companies. It is like if a beer company showed up and said: “Hey everyone! We should start beer maxxing! It’s good for your health!” Obviously, two things are happening there: the beer company makes more money when you drink more beer, and beer is absolutely not healthy — even if it is delicious. Same energy with token maxxing.
Now, here is where I am going with this: AI companies are burning cash like crazy. Running these frontier models is expensive, and that is before we even talk about training them. So what is the easiest way to make users spend more? Convince everyone that spending more tokens is actually the smart thing to do. The loop is beautiful, honestly: you start token maxxing, you spend more tokens, you spend more money, you build the habit, and then one day OpenAI or Anthropic raises prices and you barely flinch. At that point, token maxxing is not a tactic anymore. It is your workflow. And if your competitors are doing it too, then congrats, now it feels mandatory.
And to be clear, I don’t really blame the AI companies. This is their business. They need to sell more, grow faster, survive, raise, train, repeat. That is the game. But at some point, token maxxing stops being exciting and starts becoming unsustainable. Not unsustainable for OpenAI or Anthropic, by the way. Unsustainable for the companies paying the bill.
Companies are already waking up to the fact that spending around 10k a month per employee on AI is insane. Suddenly, one employee is not just one salary, one laptop, and some SaaS subscriptions. Now every employee can become a tiny GPU-powered expense monster. And that is when the most boring but important question comes back: where is the ROI? Are we actually getting value from all this AI spending, or are we just lighting tokens on fire and calling it innovation?
Ladies and gentlemen, I present to you the next metric: ROI maxxing. This is what I think will become the next big “performance metric” used to justify AI spending. And unlike most performance metrics, this one is going to be messy, because it changes from job to job, team to team, and workflow to workflow. Companies will start asking things like:
- Did AI actually make this task better, or did it just make it more expensive?
- Could this have been done with a cheaper model, or did we really need Fable 5 on ultra?
- Why is employee A getting better results while spending half of what employee B spends?
These are the kinds of questions companies will start tracking. And honestly, we might not even need to build these metrics ourselves, because the labs will probably package them for us and sell them back as a feature.
Because here is the thing: price per token might not be the best business model for AI labs anymore. It is not cleanly scalable, and customers are already struggling to justify the bill. So what happens next? The labs stop saying, “Pay us for tokens,” and start saying, “Pay us based on the value we create.” In other words: they measure the ROI you get from AI, then take a percentage of the upside. Suddenly, they are not a vendor anymore. They are your AI business partner.
I strongly believe this is coming in 2027. More and more AI companies will start positioning themselves as growth partners, not token providers. Companies will not just buy usage anymore. They will buy promised revenue lift, productivity gains, margin expansion — whatever number sounds best in the board deck — and the AI lab will take a cut. Anyway, that is the future I see. Pricing models always evolve, and price per token might already be starting to feel very 2022.
The ROI maxxing curve