There is a pattern that plays out in enterprise software every decade or so. A vendor offers generous, subsidized pricing to get companies dependent on a product, then adjusts the commercial terms once switching costs are high. What is happening in AI right now fits that pattern almost exactly.
For two years, AI companies competed aggressively on price. Flat monthly fees, bundled access, discounted pilots. The message was: use as much as you want, get comfortable, build processes around this. Now the meter is running.
Anthropic, OpenAI, and GitHub have all moved from flat subscriptions to usage-based billing, where every piece of text sent to the AI and every word it produces costs money. A global survey of over 2,000 senior executives found that 29 percent cannot explain where their AI costs are coming from. Nearly half have already slowed or restructured their AI projects because the bills exceeded what they expected.
What makes the cost problem hard to predict is a shift in how AI tools are being used. Early tools were simple: ask a question, get an answer. The newer category, where the AI works through a multi-step task on its own, reading files, writing code, checking its own work, and repeating, consumes vastly more per task. Industry data consistently shows these multi-step systems require 5 to 30 times more usage than a basic chatbot. Most company budgets were built on the old assumptions.
The numbers that have started surfacing illustrate just how badly the math can break. One company reportedly spent $500 million in a single month on AI fees after failing to set any limits on employee access. A fintech startup employee burned through $81,000 in a single week building a game with an AI coding tool. These are outliers, but the direction of travel is the same for everyone: costs are rising faster than the productivity gains are coming in.
Gartner published a forecast in June 2026 that should concern anyone making decisions about AI-heavy workflows: by 2028, the cost of AI coding tools per developer will exceed the average global developer salary. That threshold has already been crossed in markets where salaries are lower, like India. The cost of using these tools is the same everywhere in the world, but the salaries of the people using them are not. A company paying developers $15,000 a year may already be spending more than that on AI fees for those same developers.
The vendors themselves do not make this easy to manage. Gartner noted that most AI coding tool providers offer poor visibility into how usage is calculated and billed, and very few have built-in controls to help companies limit spending. The discipline to control costs is falling entirely on buyers who have no established playbook for it.
One response to this has come from an unexpected direction. Tejas Chopra, a senior engineer at Netflix, got a $287 bill from a single debugging session and started investigating why. He found that most of what was being sent to the AI was redundant: database structures, log files, and repeated formatting that the AI did not need and was being charged for anyway. He built a free open-source tool called Headroom that sits between your AI tools and the AI provider, stripping out that unnecessary content before it runs the meter. Reported savings range from 60 to 95 percent on the right kinds of workloads. The project has nearly 40,000 GitHub stars in five months.
Headroom works best for the specific type of usage that drives the biggest bills: automated workflows where the AI is processing large volumes of machine-generated data like logs and database outputs. It is less useful for simple conversational use. The data never leaves your own systems; the compression happens locally before anything is sent.
The broader lesson here matters for any operator, not just those running technical teams. AI pricing is not stable and is not designed to stay stable. The companies that set no spending limits and assumed costs would remain low are the ones getting the largest surprises. The straightforward thing to do is treat AI spend like any other operational cost: set limits, monitor usage, and require that any expansion be justified against measurable output. Companies that do this are, according to the same KPMG research, five times more likely to be able to demonstrate a return on their AI investment than those without cost controls in place.
The AI companies need to make money. They are spending enormous sums building the infrastructure to run these systems, and they are not profitable yet. That pressure will continue pushing prices up, or at minimum, will prevent them from falling as fast as they might otherwise. Treating AI costs as a fixed overhead, set once and forgotten, is the kind of assumption that will produce a very bad invoice.