Everyone is complaining about AI costs. Subscription limits, API bills, token overruns. You’d think the whole industry just found out electricity isn’t free. 🤷♂️
Here’s what nobody wants to say out loud: you were never paying the real price. For the past two or three years, you were riding on hundreds of billions in venture capital being burned to subsidise your access. OpenAI, Anthropic, Google… all of them were in an arms race for users and market share, and they priced accordingly. Free tiers, absurdly cheap APIs, unlimited plans at consumer prices. It was never going to last. These are companies with real infrastructure costs. The bill was always coming. It just took a while to arrive.
Now it has, and people are acting shocked.
I’m not shocked. I never was. Anyone who has been in this space long enough (and I mean before ChatGPT turned AI into a trending topic) knows that this kind of compute has always been expensive. The only thing that changed is that it briefly became cheap enough for everyone to stop thinking carefully about what they were doing with it.
The Waste Is in How People Use These Tools
The conversation about AI costs is almost entirely focused on the wrong thing. People are fixating on the price per token and the monthly subscription tier to determine which model costs less. That’s where very little of the waste actually lives.
An IDC survey of 318 senior decision-makers, published in December 2025, found that 96% of organisations deploying generative AI reported costs higher than expected, and 71% admitted they have little to no visibility into where those costs are actually coming from. Let that sit for a second.
Nearly three-quarters of the people running AI deployments at scale don’t know what they’re spending money on. That’s a discipline problem, not a pricing problem. 🫣
The same research found that token consumption and hallucination remediation (the cost of fixing bad outputs) were the top unexpected cost items. Both of those are controllable. Both come down to how you build and run your workflows, not what the provider charges per million tokens.
What the Token Pricing Actually Looks Like
It’s worth grounding this in real numbers, because the panic is sometimes based on vague anxiety rather than actual maths.
As of June 2026, Claude Haiku 4.5 costs $1.00 per million input tokens and $5.00 per million output tokens. Claude Sonnet 4.6 is $3.00 input / $15.00 output. Opus 4.8 (the flagship) is $5.00 input / $25.00 output. OpenAI‘s GPT-5.4 mini sits at $0.75 per million input tokens. Google’s Gemini 2.5 Flash-Lite is $0.10 input / $0.40 output, one of the cheapest production-grade models from any major provider.
For most business tasks (summarising a document, classifying a support ticket, drafting a product description), you are spending fractions of a cent per operation. The costs balloon when you use a flagship model for tasks a lighter model handles just as well, when you send rambling, ambiguous prompts that generate rambling, ambiguous output, and when you call an LLM for work a simple script could do for nothing.
There’s also a broader context worth knowing. Stanford HAI’s 2025 AI Index documented that the cost of querying a model at GPT-3.5 quality fell from $20.00 per million tokens in November 2022 to $0.07 by October 2024, a reduction of more than 280-fold in roughly eighteen months. The direction of travel is down. Judgement about when and how to use AI is the binding constraint now, not raw token price.
The AI Tax
I think of every AI call as having a cost, which I call the AI tax. Before triggering any AI workflow, the question I ask is simple:
Does the expected value of this output exceed the total cost of producing it?
That cost includes the tokens, yes. It also includes the time to review the output, the retries when the prompt was vague, the validation work when the model got it wrong, and the engineering time that went into building the workflow in the first place. Add all of that up, and a lot of “cheap” AI usage is anything but.
The MIT NANDA project published a study in July 2025 covering $30-40 billion in enterprise generative AI investment. Their finding: 95% of those organisations saw zero measurable return in their profit and loss. Ninety-five per cent. The problem wasn’t that AI doesn’t work. Most organisations never asked whether a given use case would actually generate more value than it consumed. They adopted AI because everyone else was, applied it everywhere they could think of, and then wondered why it wasn’t paying back.
The ones who saw returns (the 5%) were disciplined about use-case selection. They treated AI as a lever to pull when it created genuine value, not as a default tool for every problem.
That’s the only way I’ve ever approached it. If a workflow can’t be justified on ROI grounds, I don’t build it. Full stop.
Using LLMs for Everything Is an Expensive Mistake
A large language model is a probabilistic reasoning engine. For tasks with deterministic, predictable outputs, that’s a liability you’re paying for. A peer-reviewed study published in 2024 found that even with deterministic settings, LLM accuracy varied up to 15% across runs, with a best-to-worst gap reaching 70% on certain tasks.
Arithmetic, formatting, data validation, rules-based routing, lookup operations: none of these needs a language model. They need a script. A script runs identical logic at near-zero marginal cost, regardless of volume. Each LLM call scales with context and token count. If you’re running ten thousand records through a model to apply a conditional formatting rule, you’re paying flagship contractor rates to flip a light switch.
The right architecture uses deterministic code for stable, high-volume, predictable paths, and reserves probabilistic reasoning for tasks where language understanding and judgement actually matter. That’s where LLMs earn their cost.
On model selection: research from LMSYS RouteLLM, peer-reviewed and presented at ICLR 2025, found that routing requests between models rather than defaulting everything to a frontier model cut costs by 45-85% while retaining around 95% of output quality. A trained router sent queries to GPT-4 only 14-26% of the time, routing the rest to cheaper models at near-identical quality. For anyone running significant volume through APIs, routing should be one of the first things you implement.
Prompt caching (where repeated context such as system prompts, background documents and instructions is stored and reused) drops the cost of cached input tokens by around 90% on both Anthropic and Google’s platforms. Batch processing non-urgent tasks cuts costs by 50% across all major providers. Stack those two together for the right workloads, and your effective cost becomes a fraction of the headline rate.
The Cheaper Tool Trap
There’s a tempting logic that says: use a cheaper or free model, spend less. Sometimes that’s true. Often it isn’t.
A randomised controlled trial run by METR, published in July 2025 (updated version here), put sixteen experienced developers on real coding tasks using early-2025 AI tools. They took 19% longer than without AI. They predicted going in that they’d be 24% faster. Afterwards, they still believed they had been 20% faster, even as the data showed the opposite.
The study has caveats: it’s a specific population, a specific toolset, a specific point in time. But the principle it illustrates is real. A weaker tool that requires more correction, more validation, more rework doesn’t save you money. It shifts the cost from your API bill to your time. For most professionals, time is the more expensive resource.
Time to market is a cost. Every additional hour you spend coaxing a weaker model into producing usable output is an hour you didn’t spend on something that generates revenue. Iteration speed matters. Shipping speed matters. If a better model gets you there in two passes and a cheaper model takes six, the maths often favours the better model, especially when you account for the value of what you’re building.
Do the actual maths before assuming the cheaper option is the economical one.
What Discipline Actually Looks Like in Practice
Before you build any AI workflow, you should be able to answer three questions.
- Does this task really require probabilistic reasoning, or can deterministic logic handle it?
- Which model tier is sufficient for the output quality this task requires?
- Does the expected value of the output exceed the full cost (tokens, time, validation, maintenance) of producing it?
If the output is predictable and rule-based, use a script. If it requires language reasoning but the task is simple, use a lightweight model like Haiku or Flash-Lite. If the output has a clearly defined format (structured JSON, a fixed-length summary, a specific list format) define that format in the prompt from the start. Controlling output length and structure is the fastest way to reduce unnecessary output tokens, which bill at the expensive rate (typically 4-6 times the input rate).
Track cost per outcome, not cost per call. If a workflow costs you €50 per month in API spend and generates €2,000 in recovered revenue or saved hours, that’s a good workflow. If it costs €50 and you can’t tell what it produces, that’s a problem.
Turn on prompt caching for repeated context. Batch anything that doesn’t need an immediate response. Measure baseline token consumption per task before optimising, so you know whether changes are actually helping.
Gartner surveyed 353 technology and data leaders at the end of 2025 and found only 39% were confident their AI investments would have a positive financial impact. For the other 61%, that confidence gap is largely a measurement problem. You cannot manage what you don’t measure.
The Uncomfortable Conclusion
The AI cost panic is, at least in part, a consequence of sloppy adoption. A lot of organisations and individuals loaded up on AI tools during the investor-subsidised era without developing any discipline about how they used them. Now that real pricing is starting to bite, they’re experiencing the predictable result.
The path forward is to do what should have been done from the start: define what you want to achieve, pick the right tool for that specific task, write clear prompts that define the output you need, and measure whether the result is worth the cost.
I’ve been building on this basis since before most people knew what a token was. A €100-per-month account that produces work you can sell for thousands is not expensive. It’s one of the better returns available in digital marketing right now. The constraint was always the thinking you bring to it, not the price on the invoice.
Leave a Reply