Why Your AI Bill Keeps Growing When Token Prices Are Falling

The $20 plan that couldn't stay at $20

Cursor, one of the fastest-growing AI coding tools in the world, built its early success on a promise developers loved: a simple monthly price. For $20, a Pro user received 500 requests to the most capable AI models. One number, no surprises.

Then the way people used the product changed. Developers moved to heavier, more capable models and began handing the AI whole multi-step tasks instead of single questions. A “request” stopped being a predictable unit. One might cost a fraction of a cent; another, an agent working through a large codebase with a frontier model, could cost many times more. A flat price was sitting on top of a cost that was anything but flat.

That was the external problem. The internal one was harder: users trusted the simplicity, and any change would feel like a broken promise. Underneath it sat a principle most customers share — people deserve to know what they will pay before they pay it.

In June 2025 Cursor switched its Pro plan to usage-based billing tied to what each model actually costs to run. The rollout went badly. Many users hadn't understood that “unlimited” applied only to Cursor's automatic model-routing mode, not to every model, and some received charges they didn't expect. On 4 July, the company published an apology, said it had not communicated the change clearly, and offered refunds for unexpected usage between 16 June and 4 July.

Cursor is not a cautionary tale about a careless company. It is one of the most sophisticated AI businesses there is, and it still discovered its true cost structure in public. That is the point. If the specialists can be surprised by how AI costs behave, every organisation putting AI into its own software should assume it will be too — unless the cost is designed in from the start rather than discovered on the invoice.

The paradox, in numbers

The price of an AI token — the unit models use to measure the text they read and write — has fallen dramatically. Yet total spending keeps climbing. A February 2026 survey of 500 finance leaders by DoiT and Sapio Research found 79% of enterprises had experienced AI cost overruns in the previous twelve months. Strikingly, organisations that rated themselves most mature at cloud cost management were even more likely to overrun.

The FinOps Foundation's 2026 survey found that 98% of FinOps teams now manage AI spending, up from 31% two years earlier, and that inference — running models in live use, rather than training them — accounts for the large majority of AI cost. A September 2026 Accenture survey of 750 executives put the dynamic in a single line: companies expect token volumes to grow far faster than per-token prices fall, so bills rise even as unit prices drop.

Economists will recognise the pattern. In the nineteenth century William Stanley Jevons observed that more efficient steam engines increased Britain's coal consumption rather than reducing it. When something becomes cheaper to use, we use far more of it. AI is following the same curve.

Five reasons your bill grows

Understanding why costs rise is the first step to controlling them. Almost every overrun traces back to one or more of these.

1. More use. Work that was too expensive to automate last year becomes worth automating this year. Usage spreads across departments, often without anyone owning the total.
2. Agents multiply calls. A chatbot answers one question with one model call. An agent plans, calls tools, retrieves documents, checks its work and retries — often ten or twenty calls for a single task, each carrying the growing history of the task so far.
3. Context inflation. Retrieval systems that search your documents send passages to the model with every question. Send twenty pages when three would do, and you pay for twenty pages every time.
4. Hidden reasoning. Many advanced models “think” before answering, generating internal reasoning that is billed as output even though you never see it. For simple tasks, that thinking is pure cost.
5. The wrong model for the job. Sending routine classification or formatting work to the most powerful available model is like couriering a postcard by private jet. It works; it just costs far more than it should.

How AxON approaches this

Cost-to-run is one of the reasons we built AxON, our AI delivery platform, the way we did. The same discipline that keeps our own AI-assisted delivery efficient is built into the AI features we create for clients: each task is routed to the smallest model that does it well, repeated answers are reused rather than regenerated, and only the context a task actually needs is sent. AxON is model-agnostic, so those choices are made on fit, cost and data rules rather than loyalty to one provider. And before any client signs, we give them a written estimate of what the AI features will cost to run each month at their real user numbers.

A worked example

Consider a common request: an assistant that lets 200 staff ask questions of the company's SOPs and policies. Assume each person asks ten questions a working day, across 22 working days. That is 44,000 questions a month. The figures below are illustrative, to show how design choices move the bill, not a quote for any particular model.

The naive design sends around twenty pages of retrieved text — roughly 15,000 tokens — plus the question to a frontier model every time, and receives a 500-token answer. That is about 680 million tokens a month, all at frontier-model prices.

The designed version makes three changes. It limits retrieval to the three or four most relevant passages, around 3,000 tokens. It caches answers to the questions that recur — in policy assistants, often a third or more of the total. And it routes the straightforward questions, perhaps 70%, to a smaller model, keeping the frontier model for the complex ones. For the arithmetic, assume a third of questions are answered from cache and the smaller model costs a tenth as much per token — both conservative in many real deployments.

The result is roughly 108 million tokens a month instead of 680 million, and because most of those go to a cheaper model, the effective cost falls by around 90% or more against the naive design. The answers the staff see are no worse. The only thing that changed was the design.

That is why cost control belongs in the architecture, not in a finance review six months after launch.

Why the forecast was wrong in the first place

Most AI budgets are built the way software budgets always were: a number of seats multiplied by a monthly price. That model worked for subscriptions because cost did not change with effort. AI does not behave like that. Its cost scales with how much work it is given, how that work is structured, and which model does it. A budget written as a flat figure is really a guess about behaviour nobody has measured yet.

The fix is to budget AI as a rate rather than a total: cost per answer, per document, per completed workflow, multiplied by expected volume, with a range for growth. That forces the right conversations early. What is one completed task worth to the business? What happens if adoption is twice the plan? Which features would we switch to a cheaper model first if spending ran ahead of value?

It also settles a question many organisations leave open until something goes wrong: who owns the AI bill. When the cost sits in a central IT line, nobody in the business sees the price of their own usage. When each feature's cost is attributed to the team that uses it, behaviour changes quickly and the conversation moves from total spend to value per task.

Choosing models without bias

It is tempting to treat this as a question of which provider is cheapest. It rarely is. The global market offers genuine choice: closed frontier models from OpenAI, Anthropic, Google and xAI; strong European options from Mistral; open-weight families including Meta's Llama, DeepSeek and Alibaba's Qwen; and regional models such as Falcon, from the UAE's Technology Innovation Institute, and Jais, developed for Arabic. Prices and rankings shift every few months.

Three principles hold regardless of provider.

1. Compare cost per completed task, not price per token. A cheaper model that needs three attempts, or produces answers someone must correct, is not cheaper.
2. Self-hosting is not free. Running an open-weight model in your own environment can make sense for data residency — a real consideration in the EU and the Gulf — but idle hardware still costs money. Industry analyses suggest GPUs in live AI use often sit well below full utilisation.
3. Keep the ability to switch. A system built to route across several models can move to a cheaper or better option when one appears. A system hard-wired to one provider cannot.

Six decisions to make before you build

If you are commissioning software that includes AI, these six decisions determine most of what it will cost to run. Agree them in the design, not after launch.

1. A cost ceiling per task. Decide what an acceptable cost is for one answer, one document, one completed workflow, and design to it.
2. Routing rules. Which tasks go to which class of model, and who can change them.
3. Retrieval limits. How much of your content is sent with each request.
4. Caching. Which answers can be reused, and for how long before they must be refreshed.
5. Agent limits. A maximum number of steps and retries per task, so a stuck agent cannot loop indefinitely.
6. Attribution. Cost tracked per feature and per team, so the bill can be explained line by line.

What to ask whoever builds it

If a vendor is building AI into your software, ask for a written monthly running-cost estimate at your expected user numbers, the assumptions behind it, and which of the six decisions above the design already makes. Ask what happens to the bill if usage doubles. A vendor who can answer clearly has designed for cost. A vendor who says it depends on usage and leaves it there is handing the risk to you.

The bottom line

Your AI bill grows because usage grows faster than prices fall — and because agents, oversized context, hidden reasoning and poorly chosen models multiply the tokens behind every task. None of this is inevitable. The organisations that keep AI costs predictable make a handful of design decisions early, measure cost per completed task, and keep the freedom to choose between models.

CTA: If you are planning AI features and want to know what they will cost before you commit, we will map one use case and give you the monthly figure in writing, with every assumption shown.

Frequently asked questions

Why are AI costs rising if token prices are falling?

Because consumption is rising faster. Agents make many model calls per task, retrieval sends large amounts of text, and reasoning models generate hidden output — so tokens per task grow even as price per token falls.

How can I reduce LLM costs without losing quality?

Route simple tasks to smaller models, limit how much context is sent, cache repeated answers, cap agent steps, and measure cost per completed task rather than per token.

Is self-hosting an open-weight model cheaper?

Sometimes. It can help with data residency and high, steady volumes, but hardware costs continue when usage is low. Compare total cost per task, including infrastructure and operations.