Cheap per token, costly per task: The AI bill trap
Monday morning. Your coffee sits beside the keyboard, and overnight, your AI test has become a live process. One bill shows a cost per million tokens. Another separates inputs from responses. Which figure matters for your business? The uncomfortable answer: the token price alone tells you little about what a task will ultimately cost. Especially when an AI agent takes several steps. How do you keep the bill under control before a useful trial turns into a running expense?
Two counters make up the bill
A token is the unit many AI providers use to charge for usage. Two amounts matter for your bill: the text and data you send to the model, and the response it sends back. Providers call these input and output. If you look only at a combined price per million tokens, you miss the difference.
Take a familiar office task: you ask AI to summarise a long customer conversation and draft a reply. The conversation counts as input, and the drafted reply counts as output. The two can have different prices. So whether the task stays cheap also depends on how much material you feed in and how detailed the model's response is.
That can make token prices awkward to work with at first. The good news is you do not need to count every token by hand. You do need a sense of which tasks create a lot of input, which call for long answers, and which model is doing the work.
A million tokens, very different bills
The range becomes clearer when you look at input and output together. VentureBeat lists GPT-5.6 Luna at US$0.20 per million input tokens and US$1.20 per million output tokens. If you are charged for one million of each, the total is US$1.40.
The same article lists GPT-5.6 Sol Standard at US$5 per million input tokens and US$30 per million output tokens. At one million of each, that comes to US$35. The difference shows why a low price for a model does not tell you what a workflow will cost. What matters too is how many tokens the workflow actually uses.
These are US dollar figures from a published price comparison, not a guarantee of what your specific contract will cost. OpenRouter explicitly advises checking prices on the relevant model page before planning a budget. If your business pays through a platform or contract, check which billing model applies there too.
Cheap tokens do not mean cheap tasks
A common mistake is to assume that the cheapest model will automatically be the most economical choice. In practice, what matters is whether it completes your task reliably and how many retries or extra steps it takes. A low unit price is little help if the model repeatedly needs fixing on a routine task.
An OpenRouter benchmark illustrates the possible range for a narrowly defined task. It puts the cost of triaging 1,000 support tickets at US$0.0248 with Jev 1.13, US$0.0921 with GPT Luna, and US$2.8802 with Claude Opus. This is not a general ranking. The test ran on one day, used small example sets and used one prompt set for each model.
For a trades business, the right question might be: what does it cost to prepare an enquiry so someone in the office can deal with it quickly? For a larger company, it might be the cost of reviewing a whole ticket queue. In both cases, the cost per completed task tells you more than the price per token.
Agents can wipe out the savings
An AI agent is designed to work through a task in several steps, rather than simply answer one question. That can be useful for research, sorting or a multi-part customer process. But it can also mean the agent handles more material and triggers extra work.
Gartner forecasts that inference costs per agentic workflow will rise by more than five times by 2028. According to Gartner, routing a task to an agentic reasoning model costs at least five times as much in inference as a simple chatbot interaction. Analyst Will Sommer puts agent usage at five to 30 times the tokens or tasks for the same job.
These are forecasts and comparisons, not a bill for every business. They do show why a low price per token is no guarantee of predictable costs. If you let an agent run without supervision, decide early what will stop it if a task takes longer than expected.
What it means for your business
You do not need a large AI budget to get started. Pick a task that comes around regularly, such as sorting support requests, summarising incoming documents or preparing meeting notes. Write down what the AI receives, what result it should produce and how often the task comes up.
Then check the cost per completed task. Include input and output, as well as any extra model calls if your process has several steps. A simple test using a limited set of suitable real examples gives you a better basis for planning than comparing prices on paper. Check model prices again before you budget, as they can change.
Next, set a boundary: a maximum budget, a maximum number of runs or human approval before the agent carries on. According to Gartner, 21 per cent of surveyed companies with an orchestration platform lack an automatic block that stops agents' token spending. Even a small business can make this decision in advance, rather than waiting for a surprise bill.
- Choose a recurring task and record how much effort it takes today.
- Test a low-cost model for routine work and use a more capable model only where needed.
- Set a spending limit or require human approval.
- Check regularly that the cost still makes sense for the result.
Next Monday, approval matters
Picture next Monday. A new enquiry lands in the inbox. AI sorts it, summarises the content and suggests what to do next. A member of staff checks the result before anything goes to the customer. You can also see which task incurred a cost and whether a limit was reached.
For a sole trader, that check might be a weekly look at the bill. A larger business is more likely to need clear ownership of which teams choose models and who can stop automated processes. Either way, a simple rule helps: routine tasks get a limited budget, while exceptions need a deliberate decision.
Gartner recommends making costs visible, routing requests automatically to cost-efficient models and not using frontier models by default for routine tasks. That makes the key question practical: which decisions can your AI make on its own, and when should a person take over?
Frequently asked questions
What are tokens in AI?
Tokens are the units many AI providers use to calculate usage. Billing usually separates what you send to the model, known as input, from what it returns, known as output. The amount depends on how much content a task processes and produces.
How do you calculate AI costs per million tokens?
Multiply the number of input tokens billed by the input price, then multiply the output tokens by the output price. Divide each result by one million and add them together. For example, VentureBeat lists GPT-5.6 Luna at US$0.20 per million input tokens and US$1.20 per million output tokens.
Why do AI agents often cost more than chatbots?
An agent may need several steps and model calls to complete a task. Gartner estimates that agents can use five to 30 times as many tokens or tasks as simple chatbots for the same job. Actual costs depend on the workflow, the model and the amount of content processed.
How can I control AI costs?
Start by measuring the cost per completed task, rather than looking only at the price per token. Then set a spending limit, a maximum number of runs or a requirement for human approval. You can test cost-efficient models for routine tasks and use more capable models selectively. Check current prices before planning your budget.
Interested?
Want to know what this means for your company?
Sources
- Jev vs LLM: When to Use a Decision Model Instead of Generating Text · openrouter.ai, 2026-09-19
- Claude Opus 5.5, GPT-6 Sol, GPT-6 Luna, and a new price war · Simon Willison’s Weblog, 2026-09-22
- Gartner says agent costs will rise fivefold by 2028 even as tokens get cheaper. Here is how to keep them in check · VentureBeat, 2026-10-16
- AI price wars: OpenAI cuts GPT-5.6 Luna prices by 80% as model competition shifts toward cost · VentureBeat, 2026-07-30
- ChatGPT Rate Card (Enterprise token-based pricing) | OpenAI Help Center · OpenAI Help Center, 2026-10-24
- Build more responsive apps with the Decisions API · OpenAI, 2026-10-22
This article was researched and written with AI assistance and automatically checked against the sources listed before publication. More on our AI transparency
Klarframe AI Readiness Check
How ready is your company for AI?
About ten minutes, instant result with your stage, a 90-day plan, suitable solutions and notes on the EU AI Act.
