- Essays
- 12 min read
- June 2026
How much will your AI really cost you? And will it still be there next week?
We've spent the last fortnight unable to answer either for our own business, and judging by the group chats, nor can anyone else.
By The NowComms Team
Let’s say the quiet part loud, shall we? You’ve been bamboozled. Not by a dodgy supplier or a fudged invoice, but by a tool you’re paying for, that you use every day, that you genuinely can’t get a straight answer out of when it comes to what it’s going to cost you next month. Welcome to AI pricing, where the meter runs, the bill arrives, and the explanation is roughly “it depends.”
You just wanted to write a few emails. Turn some bullet points into something readable. Maybe let the AI have a go at that brief you’ve been avoiding. Normal Tuesday stuff. And yet somehow you’re confused, staring at a usage bar that’s flatlined, feeling mildly scammed and wondering what exactly went on in there.
We’ve been there. Credits draining for no reason we could pin down, same work as the week before, except this week it apparently cost significantly more. No note. No apology. Just a number that doesn’t add up and a support page that definitely doesn’t help.
Here’s what we think every founder and marketer should actually look at, in daylight, before it looks back at you. Two things happened recently that turned a background hum of unease into something worth talking about properly. One’s about money. One’s about control. Both are a bit discombobulating. And both have a fix that doesn’t involve giving any of it up.
Story one: In May, Uber’s own executives admitted the company had spent its entire 2026 AI budget in four months. Not nudged over by a bit. The whole year’s allocation, gone by spring, mostly on AI coding tools their engineers liked far too much to put down. The COO described the moment he realised as a “head-exploding moment”: a very polite way of saying someone walked into a board meeting with some very bad news. Fortune put the spend at hundreds to a couple of thousand dollars per engineer, per month, across thousands of engineers. That is Uber. A company not exactly short of people who can count.
Story two: Last Friday evening, the US government told Anthropic (the company behind Claude) to switch off its two newest, most powerful models for anyone outside a narrow approved group. The order was broad enough that Anthropic said it had no real choice but to pull those models for every customer while it worked out how to comply. AP and Reuters both reported it. People who had quietly built part of their working day around those models logged in to find them simply gone, for a reason that had nothing to do with anything they’d done.
Two completely different stories. One about a bill that got out of hand. One about a floor that moved without warning. The same slightly vertiginous lesson underneath both. Both happening to people who were doing everything right.
The AI everyone is racing to wire into their business is harder to price, and considerably harder to keep, than the pitch lets on. Here’s the honest version: hype removed, practical fixes included, and we promise it doesn’t end with “so go back to spreadsheets.”
01
The same work, a different bill
Start with the thing you can feel in your own account, because that’s where most of us met this problem first. Usage-based AI pricing means you’re not really buying a product at a price. You’re renting a meter, and the meter runs at a speed nobody will tell you in advance.
Most of the big tools now show you a credit balance or a little usage bar. That looks like transparency, and it beats nothing. But it’s quietly answering the wrong question. It tells you what you’ve already spent. It doesn’t tell you what the next task will cost before you press go. And that, annoyingly, is the only number a business can actually plan around.
Ask how the charge is worked out and you’ll get a soft, reasonable, completely useless answer. It depends on the effort. It depends on the complexity. It depends on how many steps the model decides to take. All true. All of which means the cost is a variable wearing the costume of a price. The same request can be cheap on Monday and eye-watering on Thursday because the model fancied the scenic route, and you only find out after the meter stops.
Some people have started comparing AI costs to utility bills. ‘Treat it like electricity’, they say. You wouldn’t leave the lights on with no idea what it costs. Which is a fair point, as far as it goes.
Except, as folk who have stared at an energy bill in genuine confusion, when did any of us actually understand our electricity bills? Standing charges, unit rates, peak tariffs, the bit at the bottom that never quite adds up. Most people just pay it and hope for the best. The utility comparison works because AI is fast becoming a recurring operating cost, a proper line on the P&L rather than a fun experiment. It’s ironic that the analogy is actually more accurate than people intend it to be: like your energy bill, the AI meter runs constantly, the numbers are real, and the breakdown, if you can even find it, is written in a language designed for someone who isn’t you.
The difference is that your electricity bill, however baffling, has a unit rate. It’s opaque, but it’s fixed. AI pricing doesn’t even give you that. The unit changes depending on what you asked, how the model felt about it, and how many steps it decided to take on the way. At least with electricity you know boiling up the kettle for your cup of tea costs more than a lamp. With AI, the kettle sometimes costs the same as running the whole kitchen, but you only find out afterwards.
We watched it happen to us. Same working rhythm, same jobs, and then a stretch of days where the credits drained like someone had left a tap running. Nothing had changed. We still can’t fully explain it, and that is exactly the point. If you can’t explain last week’s bill, you’ve got no honest way to promise the next quarter’s one.
02
Uber’s head-exploding moment
This is where someone always points out, helpfully (?!) that the answer is just to budget better. Uber is the ideal reply to that, because Uber did not blow its budget by being sloppy.
What broke their forecast was the shape of the pricing, not the size of their finance team. Old software is sold by the seat. One person, one licence, one predictable line that grows as you hire. New AI tools are billed by how much they do. One engineer asking for a quick suggestion costs close to nothing. That same engineer turning a model loose on a big job for an afternoon can rack up thousands in a single sitting. Multiply that by thousands of people who’ve just discovered how good the tool is, and your neat straight line stops being a line and starts being a wall you walk into, and fast.
The detail that sticks is that nobody at Uber did anything reckless. The engineers were being productive. The tools were working. The spend was the direct result of things going well, which is the most discombobulating kind of overrun there is, because there’s no villain and no obvious moment where anyone should have shouted stop.
It’s not just Uber either. Around the same time, Meta apparently had an internal “leaderboard” problem where staff left AI agents running with no real task just to climb a usage ranking. Different cause, same punchline: when the meter runs on its own and nobody’s really watching it, the number at the bottom of the month turns into a surprise party you did not ask for.
So here’s a small, slightly uncomfortable exercise before you scale anything on AI. Try to write the line item. Put next year’s AI cost in your budget as one confident number, the way you would for rent, or payroll or software seats. You’ll find you can’t do it with a straight face. And if you can’t price it, you can’t really plan to grow on it, however much you love it.
03
Why nobody will give you a number
It’s tempting to interpret all this as the AI companies being deliberately cagey. The truth is more structural, and honestly a bit more awkward, because it won’t be fixed by anyone simply deciding to be nicer.
The first reason: genuinely autonomous work has no fixed cost. A model that chooses its own steps can’t quote you a price upfront any more than a taxi can quote a fare before it’s seen the traffic. The more capable these tools get, the more they work this way, and the foggier the bill becomes. Capability and predictability are literally walking in opposite directions, waving at each other.
The second reason is that the prices you see today might not be the actual real prices yet. A lot of the tools people are relying heavily on are reportedly being sold for less than they cost to run. Cursor, for example, has been described as operating at heavily negative margins, paying its model providers more than it collects from subscribers. That gap closes one of two ways: the friendly per-seat plan quietly becomes usage-based, or heavy users get nudged onto enterprise contracts. Either way, the comfortable monthly figure you signed up for is a sample, not a final settled price.
Now aim that at where a lot of businesses are heading: agents, automations, the dream of letting software run on its own at scale without a human hovering over every step. Brilliant dream. But if that software is billed by how much it does, and an automated process can loop, retry or quietly wander off, then you’ve built a machine that can spend with no ceiling while you sleep. So it’s quite possible you wake up to a four-figure bill from an overnight job that no-one was watching. Funnily enough, that’s the part of the pitch the excitement tends to skip past.
04
The floor moved on Friday
At least cost is the problem you can see coming, if you squint a bit. The second one gives founders a colder feeling, because you can’t budget your way around it. The harsh reality is that you don’t own the very thing you’re building on.
Last Friday made that uncomfortably concrete and real. A US government export-control directive forced Anthropic to suspend access to its two newest, most capable models, for any foreign national, including foreign nationals inside the United States, and including the company’s own foreign-national staff. The reason given was national security. But everyone lost access, while the company worked out how to comply. Other models stayed up. Those two went dark.
The AI communities turned it into memes within a day with jokes about showing your passport before running a prompt. It lands because it’s nearly true. And underneath the joke is the part that isn’t funny at all.
Think about what that meant for a business that had built on those models. You’d done nothing wrong. Your account was in good standing. Your work was completely legitimate. And on a Friday evening, the floor under part of your business was lifted by a government you’ve never met, for a reason you couldn’t have predicted and can’t argue with. What do you even do with that?
05
And it doesn’t even take a government
Here’s where the dramatic headline becomes something that could happen on an ordinary Tuesday. A government switching off a model is the rare, cinematic version. The common version is much smaller and closer to home: one automated flag, and your account is gone.
This isn’t a new problem and not one confined to AI either. It’s a platform problem that’s been quietly wrecking small businesses for years. A gym owner in Pennsylvania had her Facebook and Instagram accounts suspended: 95% of her business was gone over a violation she says was completely false. No number to call. Just automated systems that went round in circles. Nearly fifty thousand people who say they were wrongly banned have signed a petition pointing at Meta’s AI moderation. Business owners locked out of Google tools tell the same story: automated suspensions, false positives, no human to reach, not forgetting the bonus twist where appealing too many times locks you out of appealing at all.
That same pattern is now happening in AI. Spend ten minutes in any AI users’ group and you’ll find someone who logged in one morning to a locked account and a final-sounding email, with no clear reason and no obvious route back, desperately asking everyone else how on earth they get their work back. The enforcement is a black box, the stakes are your livelihood, and the appeal form responds in the same cold voice that locked you out.
When Italy temporarily switched ChatGPT off in 2023, it worked like a fire drill. The businesses that had wired their whole workflow into one tool panicked. The ones who’d stayed a little cautious, kept a fallback, kept the important bits in their own hands? Barely looked up. The ban didn’t create that gap. It just showed everyone who’d quietly assumed the tool would always be there.
06
What we’d actually do, if we were you
Good news, and we really mean it: almost none of the fix is expensive, and none of it is “stop using AI.” It’s mostly a clear afternoon and a more honest conversation with yourself. Here’s what we’re actually doing ourselves:
Treat AI as a variable cost, not a subscription
File it next to the things that vary in cost, like ad spend, shipping, and give it a buffer you’d be a little embarrassed to actually need. The teams getting caught wrote a confident flat number. The ones who are sleeping fine at night wrote a range with room at the top.
Turn the caps on. They’re already there.
Most of the famous overruns happened on tools that had spending limits sitting right there, switched off, a bit like a smoke alarm with the battery taken out. Switch them on. Set a ceiling per task and per person, get a ping when something crosses it. Tag your usage so you can see where the money actually went. The aim is not to spend less. It is to never be ambushed by the total again.
Do the maths before you scale the robot, not after
Before you let an automated process run at ten times the volume, work out what ten times the volume costs when every step is on the meter. If the honest answer is “no idea,” then that’s your answer. Test small, watch the dial, then hand things over.
Own your exits
Keep your own copies of the things that matter. Export your data while everything is calm, not in a panic at 2am. The people who survive a lockout are just the ones who already had their important stuff somewhere they control.
Do not build your core on something you cannot swap
Use the tools, lean on them, enjoy them. Just keep the load-bearing parts of your business arranged so no single model or account is irreplaceable. If one provider went dark on a Friday, how long until you were running again on another? If the honest answer is “weeks,” that’s a dependency worth loosening now, while it is cheap and calm rather than urgent and terrifying.
Keep the human-ownable things in your own hands
The parts of your business nobody else can switch off are your judgement, your relationships, your data and your understanding of your customer. That’s the ground you actually stand on. The AI is the power tool. It’s brilliant, but it is not the floor, and the trouble starts the moment you forget the difference. Keeping an actual person in charge of the parts that matter is a recurring theme and one we’ve covered in The Pause Between Prompt and Publish.
If you’d rather not untangle all of this on your own, mapping where you’re exposed and where you are OK is a fair chunk of what our analytics and measurement work has quietly become. But the first step on this list, writing down what you’d do if a tool was taken away tomorrow, you should do today, by yourself, before you talk to anyone. Us included.
07
So here’s the deal
Don’t panic. But do read this.
None of this is “go back to doing everything by hand.” We use AI every day, it’s changed how we work, and we’re not giving it up, and neither should you.
The mistake isn’t using AI. The mistake is building something load-bearing on it while quietly treating it like a fixed, owned, predictable utility. When in reality it’s actually a variable, rented, movable one. You can absolutely build on rented ground. People do it brilliantly every day. You just have to know it’s rented, price in the fact that it might move, and never confuse the floor you’re standing on with the floor you own.
Used with your eyes open, AI is one of the best things to happen to a small, ambitious business in years. Used with your eyes shut, it’s a foundation you do not control and cannot price. Same tool. The entire difference is how clearly you are willing to look at it.
So, two questions
How much will your AI really cost you? And will it still be there next week? Nobody’s going to put either answer in writing, because in the honest cases, nobody actually knows. Budget for the unpredictable. Keep a copy of what matters. Keep a fallback. Hold your judgement, your customers, your data, the parts that are actually you, in your own hands.
Do that, and none of this needs to keep you up at night. AI goes back to being what it should be: a genuinely brilliant power tool you picked up with your eyes open, sleeves rolled, fully aware it’s a tool and not the ground.
The ground is still you. The part that holds when everything else wobbles, is still you. It was always going to be you. That’s the good news hiding inside the scary headlines, and it’s worth holding onto.
08
Questions people are asking about AI cost and risk
Is AI a scalable solution for businesses?
AI scalability is real on capability but shaky on cost and control, and that gap is the catch. Usage-based pricing means costs can rise unpredictably as you scale, as Uber found when it spent its entire 2026 AI budget in four months. Access can also change for reasons outside your control, as when US export rules forced Anthropic to disable its top models in June 2026. AI scales beautifully as a tool. It scales badly as a foundation you assume is fixed and fully yours, so the practical answer is to use it widely but build so no single model or account is irreplaceable.
Why are AI costs so hard to predict?
Because most AI tools now use usage-based pricing, where the charge depends on the work the model does rather than a fixed monthly fee. A simple request costs very little; an agent running a complex, multi-step task on its own can cost thousands in a single session. Usage bars show what you have already spent, not what the next task will cost before you run it, which is the number a budget actually needs. The more capable and autonomous the tool, the less predictable the bill.
What happened to Anthropic’s Fable 5 and Mythos 5 models?
In June 2026, the US government issued an export-control directive ordering Anthropic to suspend access to its two most advanced models, Fable 5 and Mythos 5, for all foreign nationals, citing national security concerns. According to AP and Reuters reporting, the order was broad enough that Anthropic disabled both models for every customer to comply, while its other models stayed available. It is a clear example of how access to an AI tool can change suddenly for regulatory reasons that have nothing to do with the businesses relying on it.
What happens to my business if my AI account gets suspended?
If your work lives only inside that account, a suspension can cut you off from it with little warning and slow, automated recourse, a pattern already well documented on platforms like Meta and Google, where business owners report wrongful bans, false positives and no human to call. The protection is the same on AI as anywhere else: export and keep your own copies of important work, avoid running mission-critical processes through a single account you could not quickly replace, and keep a fallback so one lockout is an inconvenience rather than a disaster.
Can a government really shut off an AI model my business depends on?
Yes, and it has happened more than once. The US restricted Anthropic’s top models in June 2026, Italy’s regulator temporarily banned ChatGPT in 2023, and the same regulator blocked DeepSeek in 2025. Bans are often temporary and tied to specific compliance issues, but the lesson is that any AI you build on sits downstream of regulators and governments you do not control. The protection is not to avoid AI, but to keep a fallback and avoid depending on a single model you could not quickly replace.
How do I reduce the risk of building my business on AI?
Treat AI cost as a variable line with a generous buffer rather than a fixed subscription. Turn on spending caps, per-task ceilings and alerts, which many tools offer but customers often leave off. Model the cost of any automation at full scale before you commit to it. Keep your own copies of important work, avoid building your core on a single model or account you cannot swap, and keep the parts that cannot be switched off by someone else, your data, judgement and customer relationships, firmly in your own hands. If you would like a second pair of eyes on where you are exposed, we are happy to take a look.
Written by the team at NowComms, a B2B marketing agency in the UK. We help founders and operators use AI with their eyes open, getting the upside without quietly betting the foundation on it. If you’d like to talk it through, or just want a friendly second pair of eyes on where your business is leaning on AI it does not control, come and say hello.
The rest of the piece
You’ve felt the problem. Here’s exactly what we’d do about it.
The second half is the practical, hopeful part: why nobody can give you a straight price, what really happened when a government switched two major models off last Friday, why a single automated flag can lock you out of your own work, and the calm things we’d do first so AI never quietly becomes a foundation you can’t price or keep. Most of them are free, and one you can do this afternoon. Drop your first name and email to read on, and we’ll unlock the rest of the hub as it grows. No pitch, no follow-up call coming.
Related service: A fractional CMO
Keep reading the hub
Get the next piece when it lands
We are writing this hub one honest piece at a time. Drop your first name and email, and we will send a quiet note when the next one is published, and unlock the full library as it grows. No pitch, no follow-up call coming.
This form is not connected yet. Please use the contact page instead.