Why Your AI Bill Went Up When the Price Didn't

Your bill went up because the meter changed, not the price. The number on the pricing page is the same one you agreed to; what moved is how many units of it a single piece of work now consumes. Vendors ship meter changes quietly — they’re not price increases, so they don’t get an email — and the invoice arrives before the explanation does.

This is the most common overspend we see in a stack audit, and it’s invisible from the pricing page. You can’t fix it by switching plans. You fix it by finding the dial.

The four dials

Almost every AI tool meters on one of four things. Know which one you’re on and you know where your money goes.

Dial	What it counts	Who bills this way
Tasks	Steps that run	Zapier
Credits	Actions, weighted by cost	Figma, Kling, Veo, Replit
Tokens	Text in and out	DeepSeek, Gemini, Grok, Devin
Seats	People with access	Figma, and most team software

The trap is that none of these dials are flat. A task isn’t always one task. A credit doesn’t buy a fixed amount of work. A token costs a different amount depending on what hour it is and how many of them you sent at once. Four real examples, all verified this month.

1. The same step, billed three times over

Zapier bills per task — every step, every run. Since June 15, 2026, an AI step inside a Zap no longer costs one task. It costs 1 on the Standard model tier, 3 on Advanced, and 5 on Premium, and every tool call inside that step is charged again at the same multiplier.

New AI steps default to Advanced.

So the Professional plan you bought as 750 tasks is 750 ordinary steps — or 250 AI ones, before tool calls. Nothing on the pricing page changed. A workflow you built in May costs three times as much in July, and the line item still reads “tasks.”

The move: open every AI step you have and ask whether it needs tools at all. The ones that don’t drop to Standard, and your task budget triples back. That audit is usually worth more than changing plans.

2. The credit that isn’t a unit of work

A credit sounds like a currency. It behaves like a variable.

On Kling, generating video at 720p runs about 6 credits per second. At 1080p with audio it’s 12. Native 4K bills at roughly 30 credits per second — five times the 720p rate. The Standard plan’s 660 monthly credits is a couple of minutes of 720p, or about 22 seconds of 4K. Same plan, same price, and the choice of a dropdown decides whether you get two minutes or twenty seconds.

Veo does the same thing in the other direction: a Lite generation in Flow costs 10 credits and a Fast one costs 20, so Google AI Pro’s 1,000 credits is either 100 videos or 50, depending on a setting most people never open.

Replit made this explicit and honest — Agent now runs in Lite, Economy, or Power modes, so you pick the cost-versus-capability trade per request instead of paying one rate for everything. Most operators leave it on the most expensive setting and never look again.

The move: find the quality dropdown before you find the pricing page. Draft at the cheap tier and spend the expensive one on the take you’ll actually publish.

3. The seat that came with a smaller bucket

Figma sells three seat types, and they don’t carry the same AI allowance. A Full seat comes with 3,000 credits a month on Professional, 3,500 on Organization, 4,250 on Enterprise. Dev seats, Collab seats, View seats, and anyone on Starter get 500 a month, regardless of plan.

Not everything meters — AI search, layer renaming, and FigJam stickies are free. But a First Draft or a Make prototype runs 20 credits. So a developer on a Dev seat has 25 generations a month, and a PM on a Collab seat has the same. Both will hit the wall and both will assume the tool is broken.

The move: buy seats by what people generate, not by their job title. Full seats for the people making things; cheap seats for the people reading them. Seat sprawl is the single most common overspend in team software, and the AI allowance is what makes it bite now instead of at renewal.

4. The token price with a clock on it

Token pricing used to be a number. Now it’s a number with conditions.

• DeepSeek moved to time-of-day pricing: rates rise during two daily peak windows — 01:00–04:00 and 06:00–10:00 UTC — and sit lower every other hour. The same batch job costs different amounts depending on when your cron fires.
• Gemini 3.7 Flash launched at $0.75 per million input tokens and $3.75 output. That’s an introductory rate with a published expiry: it doubles to $1.50 and $7.50 on January 1, 2027. Anything you cost out on today’s number is wrong next year, and the vendor told you so in advance.
• Grok 4.6 holds at $2 and $6 per million — until a prompt crosses 200K tokens, at which point xAI reprices the entire request at $4 and $12. One oversized context window doesn’t cost a little more. It costs double.
• Devin meters quota in tokens, so cheaper models stretch it further. Cognition’s Fusion router leans on exactly that: a frontier model for the hard steps, a cheap sidekick for the mechanical ones, on the plan you already pay for.

The move: batch work off-peak, cap your context deliberately, and put a calendar reminder on any introductory rate. A price with an expiry date is a decision you’ve deferred, not a price you’ve locked.

The audit, in four questions

Once a quarter, per tool:

1. Which dial am I on? Tasks, credits, tokens, or seats. If you can’t answer in one sentence, that’s the tool to open first.
2. What’s the multiplier? Find the setting that changes how much one unit of work costs — model tier, quality tier, resolution, agent mode. There is almost always one.
3. What’s the default? Vendors default to the expensive setting far more often than the cheap one. Zapier defaults AI steps to 3x. That default is a pricing decision made on your behalf.
4. What has a date on it? Introductory rates, peak windows, enforcement dates. Write them down. The changelog on every tool page here exists for this reason.

None of this is a reason to spend less on AI. It’s a reason to know what you’re buying. The operators who get burned aren’t the ones spending the most — they’re the ones who priced a workflow once, in May, and never looked at the meter again.

A clearer stack beats a bigger stack. So does a stack you can actually read the invoice for.

Next: run the stack audit to find which layer the spend belongs to, and how much to spend on AI tools for the target number.