x
AI pricing moved from flat seats to per-use meters in July 2026. Here's the 5-minute bill audit, the spend cap, and the cost-per-outcome habit to set now.
For about three years, AI cost you a flat fee. Twenty quid a month, one seat, use it as much as you like. That pricing hid something useful from you: what any individual task actually cost. It didn't matter. The bill was the bill.
July 2026 ended that.
Three things happened in a month, none of them headline news the way a model launch is. Meta moved its Business Agent — the thing answering your WhatsApp, Instagram and Messenger messages — onto token billing from 1 August, at $2 per million tokens. A typical conversation burns somewhere around 20,000 to 25,000 tokens, so roughly 4 cents a chat. OpenAI shifted workspace-agent and spreadsheet tasks onto token-based credits in early July. Microsoft put its 365 prices up on the 1st.
Read those three together and the pattern is obvious: you are no longer paying for access. You are paying for output.
There's a fourth date already on the calendar, and it's the one worth diarising: from 1 October 2026, Meta resumes charging for service messages — the free-form replies sent inside the 24-hour customer service window — which have been free since November 2024. Almost two years of "free" with an end date nobody read.
If you're a light user, metered pricing is a discount. You stop subsidising the heavy users on your tier. Four cents a conversation is genuinely cheap for something that used to need a person.
The catch is what happens when it works. Automation is the one thing in your business designed to run without you watching it. A flat fee punished you for using a tool too little; a meter punishes you for using it more than you planned. One busy week, one unusually popular post, one automation stuck in a loop overnight — and the invoice arrives before the warning does. A thousand conversations a month, at that rate, is nearer $45 than $4.
That's not an argument against automating. It's an argument for knowing the number before you scale it. If you've been retiring one boring task a week, you now have a second question to ask of each one: not just does this work? but what does one run of it cost?
Do this once, this week. Four columns, on paper if you like.
Then act on it in three moves: cancel what you never opened, put a hard spend cap on everything metered, and keep only what buys back hours you can name. Most dashboards hide the cap under Settings → Billing → Limits. Set it at roughly the hours the tool saves you times your hourly rate, halve that for month one, and raise it later on evidence rather than optimism.
If you don't yet know which hours a tool is buying back, the 10-hour leak audit is the faster starting point — you can't price a saving you haven't found.
.png)
Metered pricing quietly changes the question you should be asking. "What's the subscription?" was always a bit of a lazy metric. The useful one is cost per outcome: what does one finished client report, one drafted proposal, one answered enquiry actually cost?
This is where sloppy prompting stops being a style problem and becomes a line item. Every re-run is billed. Three vague attempts cost three times what one clear brief costs — and the clear brief usually produces better work anyway. So brief it properly up front (who it's for, what good looks like, what to leave out, how long), ask for the exact format you need, and edit the answer instead of running it again. Before any long job, one line saves you two runs: "Before writing, give me your plan in five bullets, then wait for my go-ahead."
Here's the responsible-AI angle, and it isn't a lecture. You cannot verify what you don't measure. An automation you can't price is an automation you aren't really overseeing, and an unwatched process that spends your money is the same category of risk as an unwatched process that emails your clients.
So treat a per-job agent as what it now is: a supplier. Give it a one-page contract. Scope — these tasks only. Rate — the monthly ceiling. Check — who actually reads the output. Review — a date in the diary, not "sometime". That page prevents the three quiet failures: scope creep into client work nobody approved, a bill nobody owns, and outputs nobody reads until a customer does.
And keep the human clause in. Drafts only, never decisions. No unreviewed contact with a client. You sign it, not the model — which is the whole of the trust boundary, applied to money instead of promises.
One last thing before September: if AI made you faster and you still sell time by the hour, a meter on your costs and a clock on your invoices is a bad combination. Score each tool on minutes back per pound, then re-price.
Cheap to try. Priced to scale. Plan for the second one.
Not financial advice — figures are the vendors' published rates, re-verified 2 August 2026, and your usage will differ. Check your own dashboard before you budget.
TL;DR: In July 2026 AI pricing shifted from flat per-seat subscriptions to per-use metering — Meta's Business Agent moved to $2 per million tokens (roughly 4 cents a conversation) from 1 August, OpenAI moved workspace agents onto token credits, and Microsoft 365 prices rose. For solo founders this means light use often costs less while heavy automation costs more. The three moves to make now: run a 5-minute bill audit (list every AI tool, its real last-month charge, whether it's flat or metered, anything unexplained), set a hard spend cap in each tool's billing settings, and start tracking cost per outcome rather than cost per seat — briefing clearly up front, because every re-run is billed.
Follow @aithebrief for one AI move a day — practical, vetted, no hype.
