Most AI tools tell you what a month costs. Almost none tell you what a job costs.
We know what ours cost because we keep the receipts. By 11 August we had run twelve ordinary small-business jobs on Else, all on the Smart engine, and put every one online with its receipt and a link to the run: a bakery's cash flow spreadsheet, a plumber's quote templates, a pitch deck, a lesson plan. The twelve together came to $2.36 on Smart. The middle one cost $0.15 on Smart.
Put that next to the plans most people are offered. Claude Pro is $20 a month billed monthly. Manus Pro starts from $20 a month, with from 4,000 credits. Both are honest prices for what they sell. What they sell is access, and access is not the thing an office worker actually wants. They want the forecast, the quote, the deck.
So here is the position we have built the company on, and which you are free to reject: for occasional office work, a monthly plan charges you for the possibility of work, and a job price charges you for the work, so the job price is the fairer deal.
A subscription is a bet on how busy you will be
The arithmetic is short enough to do on a napkin. At the median receipt above, $0.15 on Smart, a $20 monthly fee would buy 134 jobs. At the most expensive receipt in the set, the Kyoto itinerary at $0.45 on Smart, it would buy 44. It would buy all twelve jobs eight times over, with about a dollar to spare.
That cuts both ways, and the first concession belongs here. If you run more than 134 median-sized jobs a month, which is about six every working day across 22 working days, a flat plan is probably cheaper for you, if its usage limits let you get that far, and simpler too. Subscriptions are not a trick. They are a fair bet for people who know they will use the thing constantly.
Most of the people we build for are not in that position. A bookkeeper needs an invoice template once. A shop owner needs a cash flow forecast at the start of the quarter and a quote template now and then. A teacher needs a six-week plan when the term changes. For them a monthly plan fixes the price while the use stays irregular, and the gap between the two is paid by the customer.
We would rather not be paid for that gap. If someone runs twelve jobs with us in a month, they should pay for twelve jobs.
Twelve receipts show a spread, and the spread is the point
A fair objection to all of this is that per-job prices are too noisy to plan around. Our receipts say the noise is real but readable.
Across the twelve jobs, the cheapest was the fractions lesson plan at $0.03 on Smart, which needed no file at all. The most expensive was the Kyoto itinerary at $0.45 on Smart, delivered as a PDF. That is a spread of about sixteen times. The median, $0.15 on Smart, sits closer to the bottom than the top, and the second most expensive, the bakery cash flow spreadsheet at $0.375 on Smart, is the 90th-percentile job in a set this small.
So "it depends" turns out to mean something specific. It depends on what you asked for, and the receipt tells you how much. It does not depend on how many days are left in your billing cycle.
The twelve were not all clean, and the receipts show that too. The scam-text warning sheet, $0.14 on Smart, was asked to fit one page and came out as two. The bookkeeping invoice template, $0.21 on Smart, is the second attempt. The first attempt produced no file, was billed anyway, and cost more than the second, so the real price of that job was more than $0.41 on Smart. Two of twelve is not a failure rate you can generalise from. It is a reminder that a job price includes the jobs that go wrong.
Credits move the price to a place you cannot see
There is a middle path that many AI products have taken: credits. You pay a monthly fee, receive a number of credits, and each task spends some of them. It looks like a job price. In practice it puts an exchange rate between you and the price, and the seller sets that rate.
Manus is unusually candid about what this means. Its help centre says that "at present, Manus does not possess the capability to autonomously judge or regulate the consumption of credits", and describes a pre-task estimate as something users want and it anticipates adding "in the near future". Genspark's guide says plainly that "credits measure computational effort". Neither is hiding anything. The trouble is structural: a credit is a unit the buyer cannot convert into money without doing the seller's arithmetic, and for most tasks the seller does not publish that arithmetic.
A plumber who fixes a dripping tap does not bill you in plumbing points. You get a quote, then an invoice in your own currency, and if the invoice is wrong you can see exactly where. AI work should be billed the way trades have billed for a century: in money, per job, itemised.
Predictability is the best case for a subscription, and it is partly right
The argument for a monthly plan goes something like this. I can put $20 in a budget line and stop thinking about it. A meter makes me hesitate before every request. My finance person wants one number, not forty. People who hold that view are describing a real cost. Uncertainty is tiring, and a tool you are nervous about using is a tool you use badly.
We concede the first half. A flat fee is easier to plan, and nothing we do makes a per-job bill as boring as a direct debit.
But look closely at what the person actually wants. They do not want to pay $20 whatever happens. They want to know the most they can lose. That is a ceiling, and a ceiling can sit on top of a job price just as well as a flat fee can. Simon Willison made the general case on 3 October 2026, writing that most businesses and individuals "would prefer errors to a surprise $10,000+ bill", and hard caps should be on by default.
Work you hand off to Else carries a spend limit of $5 unless you set a lower one first. It is imperfect: it is checked after money is spent, so a run stops a little past your number, then asks before going on. A brake, not a wall.
A job price means you pay for the jobs that fail
This is the objection that hurts most, so here is the worst receipt we have.
On 3 October we re-ran one of our own July examples to check it still worked. The engine was set to Auto, and Auto picked Max. The job did produce its file. Then the step that hands the file back to you failed to recognise it, and tried again, and again, 21 times. The run stopped when it hit a $5 default limit, with a final bill of $5.08 on Max. Of that bill, $4.60 on Max, or 90.6%, went on the loop. In July the same request had come back for $0.51 on Auto. Against the gallery, $5.08 on Max is the price of 34 median Smart jobs.
It was our account, so we paid it. A stranger trying Else for the first time would have paid it too, and that is exactly the first experience that turns someone off a product for good.
We changed two things that day. Auto stopped sending every job that works with files straight to Max; a receipts-to-spreadsheet re-run on Auto then went to Smart and cost $0.36 on Smart, though it mislabelled a tax line we fixed later that day. And the hand-back step now gives up after two failed attempts on the same file and says so. The first change is checked against real runs; the second has had few chances to fire, so we treat it as new, not proven.
None of this comes with a refund. On Else today a failed attempt is billed for the work that ran, and we will not pretend otherwise. What a job price does is make that failure visible as a line on a receipt. On a subscription the same failure still costs you, as a slice of your allowance; Genspark's guide says outright that "each failed retry costs credits". We would rather show you the $5.08 on Max and what we did about it than let it disappear into a monthly total.
Recurring revenue is the argument we feel most
The best case for subscriptions is not about customers at all. It is that predictable revenue lets a company plan, hire, and keep improving the product you depend on. A business paid by the job lives closer to the bone.
That is true, and it costs us directly. A customer who runs the twelve jobs above pays us $2.36 on Smart for them. The same customer on a $20 plan would have paid $20 whether they used it or not. We give up that difference on purpose, and it limits how much we can spend to find new customers. If it turns out that people who pay by the job cannot sustain a company, we will have been wrong in the most expensive way available.
We are taking that risk because the alternative quietly depends on customers not using what they paid for. A plan that profits from unused months is paid, in effect, to be forgotten. A business paid by the job only earns when a job runs, and every failure lands on a receipt with our name on it. That is the incentive we want to work under.
A price per job changes the question you ask
The last argument is not about money. When a job has a price, the useful question changes from "am I getting my money's worth this month?" to "is this job worth about fifteen cents on Smart, the median of our twelve receipts?" Often the answer is yes. Sometimes it is no, because the task takes two minutes by hand, and you should do it by hand. A subscription never prompts that judgement. It nudges you to use the tool more, to justify the fee.
We have not measured whether showing prices changes what people ask for, and we will not claim it does until we have. We only know which question we would rather our customers asked.
If you want to check any of this, every one of the twelve receipts is linked from our AI task costs table, with the data available to download.
Method note. The twelve receipts are every gallery job with a public replay as of 2 October 2026: twelve jobs, one published attempt each, all on Smart, chosen by us to show the product rather than sampled at random. The receipts carry no timestamps; each job was added to the gallery on 10 or 11 August 2026 and its total has not changed since. The failed first attempt of the invoice template is not among the twelve, and its amount is not published. Exact totals: $2.361872 on Smart for all twelve, median $0.148155, 90th percentile $0.375, maximum $0.452333, all on Smart; the October runs were $5.0796 on Max and $0.359005 on Smart. Median, 90th percentile (nearest rank: the 11th of 12) and maximum were computed from the twelve totals; the $20 comparisons divide $20 by those figures. The $5.0796 on Max run and the $0.359005 on Smart re-run took place on 3 October 2026 on our own account. Competitor prices were read on 5 October 2026 from the linked pages; we left out any plan whose price page we could not read.
A month is a unit of time, and nobody hires help for a unit of time when what they need is a spreadsheet.