The dangers of token usage billing

Imagine the scenario: your engineering team is more motivated than ever and delivery times have dropped by 30% thanks to the integration of AI agents across your workflows. Life is great — until you open your email and see a notification from your favorite AI provider:

"Update to our terms of service: We are transitioning to a token-based credit model to ensure service sustainability."

What used to be a flat $20 a month per developer is now a variable cost that depends on how many loops the model decides to run to solve a single bug. That morning coffee you were enjoying just turned very bitter. You suddenly realize that by optimizing your operations you also set a trap; you’ve become imprisoned by a tool, the price of which is now decided by someone else.

From subscription to consumption

While this is a story that has exploded in recent weeks, it really started when we stopped using AI as a dictionary and began using it as an employee. The original Copilot was predictable: a flat monthly fee, some magical autocomplete here and there. But 2026 is the year of autonomous agents.

An agent is not a chat box. An agent receives an order, reads 50 context files, tries to run a test, fails, re-reads the context, invents a solution, discards it and starts all over again. Each iteration can burn through hundreds of thousands of tokens. Under a pay-per-use model, a single refactoring task can cost more than the server hosting the application.

We used to rely on third-party tools to make our lives easier, and we were happy to pay a monthly subscription for that service (the SaaS model). Now, however, we’ve transitioned from renting software to buying raw compute capacity. We’re paying for something else to run our processes and autonomously steer the business, simply because we assume it’s faster and more effective than doing it ‘manually’.

But is this actually true? And, more importantly, can companies afford it?

When AI bills you for thinking

Under the pay-per-use model that’s taking over, development costs are no longer a flat monthly fee. In fact, they feel like a bit of a roller coaster. Let’s review three real-world scenarios:

• Consumer one: The solo developer. A developer uses a frontier model on demand for their day-to-day work. This includes every prompt, every pull request and every process that requires planning , such as brainstorming for an upcoming ADR (architecture decision record) or writing documentation. Here, consumption is manageable. The company smiles and nods, thinking it’s cheaper than a traditional software license.

• Consumer two: Automated processes. You decide to automate code reviews, documentation generation or any other tedious process that eats up valuable time. Every time someone pushes code, an agent reads the repository context, analyzes the changes and proposes improvements. If you have 10 developers pushing three PRs a day, by the end of the month we could be talking about millions of tokens, depending on the size of the codebase and the number of reviews. And that's assuming the agent doesn't mess up and force a retry.

• Consumer three: Extensive autonomous processes. This is where things get downright bloody: you automate support ticket triaging and real-time log analysis, delegating critical decisions to a high-tier LLM. You have agents running 24/7, continuously feeding gigabytes of data into the context window with zero guardrails. Suddenly, token consumption skyrockets exponentially; the trap is that your workflow is now so dependent on these processes that shutting off the service means halting operations entirely.

The outsourcing of thought

The true danger of vendor lock-in in the AI era isn't just token price hikes, which are dangerous enough on their own. The real danger is that you’re handing over your business to the model provider.

When you delegate code comprehension and the architecture of your new systems to a proprietary, third-party model, you’re essentially transferring your company’s most valuable asset: its knowledge.

If your developers stop navigating the codebase because the agent takes care of it, you face a massive risk. The day the provider decides to close your account, change its terms or simply price the model out of reach, you’ll be left with a company where nobody actually knows what business rules apply to the system. You'll have thousands of lines of code no one on your engineering team truly understands, because the 'brain' that connected them lived on a server in San Francisco.

We’re at risk of creating a generation of companies that are empty shells. They’re capable of executing at lightning speed, but incapable of reasoning about their own technology without asking for permission (and paying the toll) to an external API.

This scenario isn’t new. In the past, database or cloud lock-in was an often expensive migration headache. Now, though, knowledge lock-in poses an existential threat to businesses.

The gradual sabotage of AI

If economic costs and knowledge loss weren't enough to worry you, there’s a third variable to consider: model degradation.

Let me introduce you to "AI poisoning" — the gradual unintentional sabotage of AI. Frontier models are trained on whatever they find on the internet, but the internet is rapidly turning into a landfill of garbage AI-generated content. While this happens across all industries, in software development inaccuracies are uniquely destructive.

If OpenAI or Anthropic models start feeding on mediocre code, obsolete security patterns or hallucinated business logic generated by their own previous versions, the quality of the responses you’re buying at premium prices will begin to tank. Trusting an external API blindly means betting on a brain that is slowly losing its sanity by consuming its own digital waste. The 'black box' tool isn’t only expensive and a jailer for your company’s knowledge; it’s unreliable.

Reclaiming sovereignty

So, what’s the alternative? Going back to manual coding? Far from it. The solution isn’t to abandon AI, but to transition from being mere consumers to becoming owners.

Here are some practical steps worth considering:

• The bet on open source. Models like Llama 4, Kimi or Mistral have proven the gap with frontier models is closing fast. 

• Local and specialized models. Instead of using one giant model for everything, companies are deploying smaller models on their own infrastructure. They are faster, infinitely cheaper in the long run (electricity cost vs. token cost), and above all, completely private.

• Fine-tuning with your own preferences. The true competitive advantage in 2026 is training or fine-tuning an open-weight model with your own clean codebases and actual documentation. A model that understands your specific patterns and business requirements is far more valuable than a generic high-end model that has been poisoned by millions of throwaway sandbox repositories.

• Agnostic IDEs and abstraction layers. Don't tie yourself to an IDE that forces you into their proprietary model. Use tools that allow you to swap your LLM provider with ease.

Control the model; retain autonomy

AI should be a tool that empowers developers, not one that replaces them or holds them hostage. The flat-rate era might be at an end, and with it, the era of corporate naivety.

The companies that survive this decade won’t be those that are able to buy the most tokens, but those that have the courage to reclaim technical sovereignty and understand their system's knowledge is the one thing that can never be outsourced.