AI tools have never been cheaper to build or buy. Token prices from Anthropic, OpenAI and others are falling fast - Chinese competitors are undercutting US providers by nearly nine to one on some workloads, and the pressure is pushing Western labs to cut rates too.
Bills are rising anyway.
The answer is agents. If you are running agents, or planning to - last week’s piece on the shift from chat to agents lays out why you should be - you need to understand this before the invoice arrives.
The token gap
A chatbot conversation uses a handful of tokens. You ask, it answers, the exchange ends.
An agent works differently. It reads context, plans, checks its work, loops back, adjusts and then acts. A single task that would take a chatbot one exchange takes an agent five to thirty of them. Every step burns tokens. The economics that held for chat do not hold for agents.
The Economist ran the headline last month: “Companies are scrambling to curtail soaring AI costs.” Not because the price per token rose - because the number of tokens a business uses, once agents are running, is orders of magnitude higher than before.
The starkest case is Uber. After opening Claude Code access to around five thousand engineers in late 2025, the company burned through its entire 2026 AI budget by April. Four months.
What Anthropic just changed
On 15 June, Anthropic split its Claude subscriptions. Agent credits and API usage now come from a separate pool to chat. If your team uses Claude for conversation and you also run it through an automation layer or the API, you need to track two budgets, not one.
Expect OpenAI to follow. Running an agent costs the labs far more to serve than answering a question, and the billing now reflects that. Most small businesses have not prepared for this.
What I see happening
A marketing agency I work with decided earlier this year to let an AI agent handle their client reporting. Every Friday it pulled data from four platforms, drafted a report for each client, and flagged anything unusual.
The work was excellent. The time saved was real - roughly twelve hours a week that a junior account manager had spent copying numbers into slides.
The bill was a surprise. The agent was touching forty-two client reports, running multiple tool calls per client, weekly. Monthly AI spend quadrupled before anyone looked at it, because nobody had set a cap.
They kept the agent. They set a cap. And they learned the lesson without losing the value.
The two things you control
You control cost on two sides: what model you use, and how much you let it run.
Not every task needs the most capable model. The big labs all offer tiered options. Routine jobs - formatting, tagging, pulling data, drafting standard responses - work fine on a cheaper, faster tier. Reserve the frontier model for the work that requires it. Think of it the way you think about your team: not every job needs your best person.
Set spending caps before you let an agent loose. Every major platform has the controls. Use them. Treat your AI spend the way you treat your cloud hosting or your phone bill - with a ceiling, not an open tap.
If your tools do not show you token usage by task, find ones that do or build a simple log. Buying tools without understanding what they cost to run is the same mistake whether the tool is a SaaS subscription or an autonomous agent.
Why this is still worth it
Agentic AI costs more per task because it does more per task. Compare the cost to three hours of a skilled person’s time, not to a chatbot session. For most tasks, the maths work.
AI remains the most powerful lever a small business can pull, because a small team can now automate work that used to require headcount. Go in knowing what it costs.
Run your agents. Set your caps. Check your usage once a week until you know your baseline. Then spend the time you get back on the work only you can do.
What to do this month
Check whether your AI provider has split agent and chat billing. If you use Claude or ChatGPT through an API or an automation tool, find out which pool your usage draws from and whether there is now a separate limit.
Set a monthly spending cap for every agent you run. Start low and raise it once you know your baseline.
Audit which tasks your agents handle. Match each one to the cheapest model tier that gets the job done.
Look at your last invoice. If the number surprised you, audit rather than abandon. Keep the agent, cap the spend, and make the value visible.
AI thinking is getting cheaper. AI doing costs more. Build that into your decisions from the start.