On 20 July 2026, AI Business reported that executives discover the cost of their AI on the following month's bill, with no way of having predicted it. And prices move both ways: Google doubles the price of Gemini 4 Argon on 1 December 2026. How much will AI cost you if you roll it out? LLM API cost can be budgeted like electricity: you measure, you cap, and you pick the appliance for the job. That is how I run my own agents, for about 46 dollars of API usage a month.
Key takeaways
- For the same level of performance, the price per token is falling fast. But the price of frontier models does not always fall, and some are going up.
- Bills rise mainly because usage rises: an agent rereads its context, calls tools and "thinks" in billed tokens.
- Within one vendor, the smallest model costs 80 to 100 times less than the most powerful one. Matching the model to the task is the first lever.
- Caching and batch processing sharply cut the price of repeated or non-urgent tokens.
- An AI budget comes down to five habits: measure, cap, choose the model, cache, alert.
What is happening: tokens get cheaper, bills get bigger
The price of a given performance level is collapsing
In March 2025, the research institute Epoch AI measured the price needed to reach a given performance level on six benchmarks. The decline ranges from 9 to 900 times per year depending on the level. The researchers themselves note that the fastest declines are recent and may not last. Price lists show it too: at Anthropic, the top model of the 4.1 generation cost 15 dollars per million input tokens and 75 per million output tokens; the current Opus model (5.5) costs 4 and 20.
Yet bills keep rising
The AI Business report quotes several executives. Starburst's CEO says his token costs are rising "faster than we thought" because his teams use more of them, and that he has no good model to predict spending: he adjusts after the month closes. Pega's CFO calls the cost of reasoning tokens a "black box". The analytics firm Silicon Data finds that per-token prices have stayed relatively stable in 2026: bills rise because of volume, or because promotional deals come to an end.
And some prices are going back up
The Gemini API pricing page states it plainly: Gemini 4 Argon goes from 2 to 4 dollars per million input tokens, and from 10 to 20 for output, on 1 December 2026. Gemini 3.8 Flash doubles in turn on 1 January 2027. AI is also moving into business software: the Odoo 20 release notes, published in September 2026, state that all AI features now require credits, bought through Odoo's in-app purchase service (IAP).
What it means for you: an agent's LLM API cost is not a chat's
In a chat, a person asks a question and reads the answer. An agent chains calls without waiting for anyone. On every call it sends the model its instructions, the description of its tools and the task history. That volume adds up fast: at Anthropic, declaring the web browsing tool alone adds about 6,600 tokens to every request.
Then come reasoning tokens. OpenAI states that these tokens, invisible in the answer, are billed as output tokens; Google lists its output prices "including thinking tokens". And at OpenAI, as with Gemini 4 Argon, an output token costs five times as much as an input token.
Finally, the gap between models is huge. At OpenAI, the smallest GPT-6 model costs 0.10 dollars per million input tokens, the most powerful 10 dollars: a hundred times more. At Anthropic, the range runs from 0.10 to 8 dollars. Mistral's Large model is listed at 0.50 dollars for input. Sending every task to the most powerful model is like paying an engineer to sort the mail.
For an SME owner, the conclusion is simple: the question is not "how much does AI cost?" but "how much does this task cost, with this model, this month?". Without measurement, there is no answer.
What I learned running my own agents
My AI agents run around the clock on my own server: they manage my ERP, my notes, my publications and part of the infrastructure. I describe the architecture in my AI agents on a server, always available.
The cost is published on my AI page: about 57 dollars a month in total, of which 7 for the server, 46 for model APIs and 4 for extras. That covers 200 to 400 requests a day, about 1.5 million tokens a day and three users. Do the maths: 45 million tokens a month for 46 dollars, roughly one dollar per million. At a high-end model's input price of 10 dollars, the same volume would cost over 400 dollars.
Three lessons have served me well.
First, every time an agent wakes up, it costs. I removed a relay that called a model on every message. Since then I apply a simple rule: no AI run is triggered by the tool itself, the AI only runs when someone asks.
Second, the model is chosen for confidentiality as much as for price. When the data requires it, a local model running on a machine I control: nothing leaves. When it is acceptable, an external API, more powerful, while controlling what I send it. Both live side by side in the same setup.
Third, a metric beats a hunch. My setup is monitored like any other server: I see what is running, what it costs and what breaks. It is the same lesson my incidents taught me: you cannot steer what you do not measure.
What to do: budgeting your agents in practice
1. Measure per task, not just in total
Give each project or agent its own API key, and record input, output and cached tokens every week. Vendor consoles provide these reports: Mistral's dashboard, for example, tracks calls and per-user activity, and breaks costs down by service. Then relate the cost to a business unit: an invoice processed, an email drafted, a ticket resolved.
2. Cap spending before rolling out
Set a monthly spending limit per project. At Anthropic, for example, a spend limit sets an organization's maximum monthly cost, and finer limits can be set per workspace. Limit response length too: OpenAI lets you cap the number of generated tokens, reasoning included.
3. Match the model to the task
Sorting, summarizing, extracting an amount: a small model is enough. Reasoning about a contract or fixing tricky code: the large model is justified. Test the task on the small model first, and only move up if the result does not hold. For sensitive data, consider a local model.
4. Cache what repeats and batch what can wait
An agent's instructions and tool descriptions are identical from one call to the next: once cached, they cost 10% of the input price at OpenAI and at Anthropic (even less on its recent models), and up to 90% less at Mistral. Jobs that can wait a few hours (archive sorting, translations, reports) go through batch processing, billed at half price by all four vendors mentioned.
5. Set alerts and review the bill every month
An alert at 50% and 80% of the monthly budget, sent to a real person. Every month, look at the three most expensive tasks and ask whether they pay off. Watch price announcements too: a model that doubles on 1 December changes your January budget.
Frequently asked questions
Subscription or API: which should I choose?
A subscription suits people who chat with AI: the price is fixed and predictable. The API suits agents and automations: you pay for usage, and you can choose the model, caching and caps. Many SMEs need both.
Is a local model cheaper?
Not always: you need a machine, electricity and maintenance time, and small local models are less capable. Its main advantage is confidentiality, since nothing leaves. For simple, repetitive tasks on sensitive data, it is often the right choice.
How do I estimate the budget before launching an agent?
Run the agent for a week on a real case, with a low cap, and measure the cost per task. Multiply by the expected monthly volume, then add a margin for announced price increases. An estimate based on a demo is almost always too low.
Going further
I design AI agents and automations with their budget built in from day one: see custom applications and automations and my guides to hosting your own AI. To find where AI will actually make you money, also read AI for small business: move from trial to use that pays off. To discuss your case, book a 20-minute call.
Sources
- As AI Spending Climbs, Enterprises Get Serious About Token Costs, Patrick Thibodeau, AI Business, 20 July 2026.
- LLM inference prices have fallen rapidly but unequally across tasks, Epoch AI, 12 March 2025.
- Pricing, Anthropic, accessed 10 November 2026.
- Rate limits (Spend limits), Anthropic, accessed 10 November 2026.
- API Pricing, OpenAI, accessed 10 November 2026.
- Reasoning models, OpenAI, accessed 10 November 2026.
- Gemini Developer API pricing, Google, accessed 10 November 2026.
- Pricing, Mistral AI, accessed 10 November 2026.
- Odoo 20 Release Notes, Odoo S.A., September 2026, accessed 10 November 2026.
- Photo: Jakub Żerdzicki / Unsplash (unsplash.com/photos/8wLZi9OhsWU), Unsplash License.
Tags: LLM API cost, AI agents, AI budget, LLM, prompt caching, SME