LinkedIn is drowning in alarmed posts about AI budgets spiralling out of control. After a year running a multi-agent platform in production, I’ll say something unpopular: a runaway bill isn’t a fact of the market. It’s the symptom of a decision nobody made.
The genre has become a ritual. A screenshot of a bill, a scary number, a call for caution about « these costs going off the rails ». Hundreds of posts a week, all built the same way.
Subscription or API: where can an AI bill actually explode?
Let’s pin down the ground first, because confusion feeds the panic. On a subscription, an explosion is impossible by design. You choose your tier, you accept or decline automatic top-up, you set the monthly limit you won’t cross.
I’ve ridden that lift at Anthropic myself: 20 euros, then 90, then 200, back down to 90, back up to 200. At every floor, a ceiling I had signed off on. Never a surprise, never a runaway, by definition. The amount is yours to decide, up or down.
That leaves a single place where a bill can genuinely drift: the API, charged by usage. And that’s precisely where almost nobody is steering.
Ask the only question that matters: what, exactly, is going to your most expensive model? The answer is nearly always the same. Everything. Classifying a ticket, tagging a piece of content, checking a length, extracting a date. Tasks an entry-level model handles perfectly go to premium, request after request, because it’s the model wired in on day one and nobody has touched it since. It isn’t AI that’s expensive. It’s the missing decision.
What do the public price lists show that alarmed posts ignore?
Price lists are public. At Anthropic, the gap between the fast model (1 dollar per million input tokens) and the reasoning model (5 dollars) runs from 1 to 5, and climbs to 1 to 10 with the very top tier. The official documentation recommends the split itself: the small model for simple tasks, the mid-tier for the bulk of production, the large one for complex reasoning. Other providers show comparable gaps.
Read that again for a second. The provider itself tells you not to send everything to its most expensive model. And it publishes two mechanisms most teams never switch on: batch processing, which halves the bill on anything that isn’t real-time, and prompt caching, which cuts the cost of repeated tokens by 90% from one call to the next.
The saving is documented, public, a config setting away. Whoever pays five times the price has chosen to pay five times the price.
How do you route models without sacrificing quality?
The editorial platform we run orchestrates 13 specialised agents across an 11-step workflow. They don’t all use the same model by default. None of this was bolted on afterwards: from day one, every stage of the pipeline knows what it’s allowed to spend.
- Entry-level — Tone of voice, SEO optimizer, schema.org, compliance legal, formatter packager
- Mid-tier — Trends & Keywords, structure outline, originality check, editor dedup, final review, linkedin comment
- Top-tier — Writer, human editorial pass
« 85% of our API calls never touch the premium model. »
The result: our costs hold steady. Not thanks to a negotiated discount, but because the routing was in the plans before the first request.
Why do so few teams actually route their models?
Let’s be fair to the teams that don’t route: it isn’t free. You have to classify the requests, maintain the mapping table, evaluate continuously. Badly calibrated routing sends hard tasks to models that fumble them, and the savings evaporate in rework. It’s systems work. Not a setting.
And that’s exactly why this goes beyond the bill. An application that routes well is an application where someone has defined what each stage demands. A controlled cost isn’t a financial performance, it’s the visible trace of an architecture that was thought through. A runaway bill tells the opposite story.
What questions put you back in control of your AI bill?
Three questions comfortably replace the next alarmed post.
- What share of your requests genuinely needs your most expensive model?
- Do your asynchronous jobs go through batch mode?
- Is your repeated context cached?
Three questions, three factual answers. And most of the time, a double-digit reduction without switching provider or giving up a single point of quality.
Next time a post tells you AI costs are uncontrollable, translate it: theirs are.
The prices quoted are from Anthropic’s public documentation, consulted in July 2026. The subscription amounts refer to my own individual account; enterprise accounts have their own rules. The editorial platform mentioned is NOMO IA, published by Fast Growth Advisors.
Sources
- Anthropic — Pricing (per-model rate cards) (consulted 17 July 2026)
- Anthropic — Prompt caching (cache reads at 0.1× input price) (consulted 17 July 2026)
- Anthropic — Message Batches API (asynchronous processing at half price) (consulted 17 July 2026)
- Anthropic — Choosing a model (matching models to task complexity) (consulted 17 July 2026)
Frequently asked questions
Can a subscription AI bill explode?
No. On a subscription you set a tier and a monthly limit, and you choose whether to enable automatic top-up. The amount is capped by a decision you signed off on. The only billing model where spend can genuinely drift is the API, charged by usage.
Why do AI API bills drift out of control?
Because every request goes to the most expensive model, including simple tasks (classification, tagging, extraction) a small model would handle perfectly. It’s the model wired in on day one and never reviewed since. The problem isn’t the price of AI, it’s the absence of routing.
How do you cut an AI API bill without switching provider?
Three levers Anthropic documents itself: route simple tasks to entry-level models (5 to 10 times cheaper than premium), use batch processing for anything asynchronous (half price), and cache repeated context (up to 90% off reused tokens). Combined, they usually produce a double-digit reduction.
What is model routing?
It’s assigning each task to the cheapest model that can handle it correctly, instead of sending everything to the most powerful one. At NOMO IA, 13 agents across an 11-step workflow use three model tiers depending on what each step demands; 85% of calls never touch the premium model.
Do batch and prompt caching degrade quality?
No. Batch processing changes neither the model nor the result: it handles non-urgent requests asynchronously at half price. Prompt caching reuses identical context from one call to the next without charging full rate again. Output quality is the same; only the cost drops.