The Economics of AI Content: Subscriptions vs Credits, and the Real Cost Per Finished Asset

Quick news, the week of June twenty-third through July second.

Biggest story. OpenAI showed up at Cannes Lions and basically declared itself an ad company. This was reported June twenty-third through the twenty-sixth. Their chief revenue officer, Denise Dresser, told brands OpenAI is "clearly in the advertising business now," and their ad chief, Dave Dugan, pitched ChatGPT users as "super intentional." The numbers they waved around: nine hundred million weekly active users, and about twenty percent of queries carrying commercial intent. ChatGPT ads started testing back in February, and they're now live in seven markets, with Brazil and Mexico next. Reportedly, and treat this as vendor talk, two and a half billion dollars in ad revenue this year, scaling toward a hundred billion by 2030. Criteo says over two thousand brands already run ChatGPT ads through its platform, and OpenAI partnered with LiveRamp for measurement. So, a brand-new paid channel living inside the assistant. Go check ad eligibility in your market.

HubSpot had one of its biggest shipping months ever. The HubSpot Agent for Microsoft 365 Copilot went live in June. It surfaces live deal, ticket, and contact data right inside Copilot, fetched in real time and never indexed into Microsoft's systems, and it's available across all hubs and tiers. They also shipped a redesigned Breeze Projects workspace with direct file uploads, wider Prospecting Agent access, and mobile Breeze meeting briefings. If you live on the Microsoft stack, turn that connector on.

ActiveCampaign, on June twenty-seventh, shipped Active Intelligence two-point-eight with "brand memory." It remembers your voice guidelines, colors, logos, and priorities across sessions. Less re-prompting.

Google Ads, June twenty-third: new AI Max reporting guidance, and Dynamic Search Ads will auto-upgrade to AI Max starting February 2027. Audit those campaigns before the forced migration. Google's also beta-testing "Ask Ad Manager," a Gemini chatbot for delivery problems.

On the agentic ad stack, all dated June twenty-seventh: Yahoo launched an Agent Network with twenty-three ad-tech partners, Warner Bros. Discovery is rebuilding on agentic AI on AWS, Amazon launched Alexa+ ads with in-ad purchases, Adobe updated GenStudio, and WPP is testing an AI buyer agent for video. Reportedly, a Salesforce study puts agentic adoption jumping from thirty-nine to sixty-six percent in a year, while a Lippincott survey found only about one in ten marketing leaders rate their tech enablement "excellent."

Assistants. Claude joined Slack as a persistent teammate on June twenty-third. Perplexity's Comet browser now lets you pick the model, with Opus 4.6 the default. And Gemini 3.5 Pro's general release slipped to July.

Your standing AI-search check. SparkToro, June 2026: sixty-eight percent of Google searches ended with no click, just two hundred seventy-six clicks per thousand reaching the open web, down from three hundred seventy-four. Zero-click hits ninety-three percent in AI Mode versus forty-three in AI Overviews. Ads now show in a quarter of AI Mode results, and Google's piloting Sponsored Stores and Direct Offers. AI Mode passed a billion monthly users. On citations, March through April: ChatGPT took sixty-three percent of B2B AI referrals, Claude nineteen, Gemini eleven, Perplexity seven, with only eleven percent of cited domains overlapping across platforms. Trackers worth watching: Profound, Superlines, and SearchSignal.

This is the last tutorial in Act One. And it's the one nobody wants to sit through, because it's about money. But it's also the episode that decides whether all the skills we've built actually pay off, or quietly drain your budget while you feel productive.

Here's the whole thing in one line. The number that matters is not cost per generation. It's the fully-loaded cost of one thing you actually shipped.

Before we go further, a framing note. Every price I'm about to say I verified the week of July second, 2026. Treat all of it as "as of mid-2026." Prices move fast in this market, so before you budget anything, open the vendor's own pricing page and check. And here's a fact that's easy to miss: some vendors keep their real pricing behind a sales call. Jasper's Business tier, Writer's Enterprise tier, the big ChatGPT and Claude enterprise deals. That opacity isn't an accident. It's a negotiating position. And the fact that a company won't publish a rate card is itself information worth naming out loud.

Okay. Let's build this up.

There are basically three ways an AI tool charges you.

The first is flat monthly seats. A subscription. You pay a fixed amount per person per month, and you get "unlimited" use within some fair-use or rate limit. The thing to understand about flat seats is what they reward and what they punish. They reward heavy iteration. Once you've paid for the seat, another draft is basically free at the margin. Generate ten headline variants, regenerate that paragraph fifteen times, it doesn't change your bill. But flat seats punish idle seats. If you buy five and only two people log in, you're lighting money on fire.

And here's the trap hiding inside flat pricing. Unlimited-seat overconfidence. You start treating that flat seat as truly free at the margin, and you forget that every "free" regeneration produces output that a human still has to read, fix, and fact-check. The token cost went to zero. The edit-hour cost did not. Hold that thought, because it's the whole episode.

The second model is credit or usage or token metering. You buy a bucket of credits, or you pay per token, per generation, per second of video. Every action draws the bucket down. What does this reward? Discipline. You think before you generate, because you can watch the meter move. But what does it punish? Exploration. That running meter creates what I'd call credit anxiety. You feel the cost of each try, so you under-iterate. You ship the first output that's merely acceptable instead of the third one that was actually good.

And that's the trap on this side. Hoarding credits. Shipping a worse asset to save pennies of credit, while you burn dollars of human edit time cleaning up the mediocre thing you settled for. Both traps, on both sides, come back to the same blind spot: counting the wrong cost.

There's a third pattern worth naming, which is that the industry is drifting toward hybrid consumption-based pricing. PYMNTS reported that AI is pushing software off flat subscriptions and toward consumption pricing across the board. And the numbers are sobering. Sixty-three percent of enterprises blew past their AI budget by thirty percent or more in the first year. Seventy-eight percent of IT leaders ate an unexpected consumption charge. Why does that matter to you, a marketer, not an enterprise buyer? Because a tool you budgeted as "flat" may quietly start metering its AI features. Canva did exactly this, and we'll get to it.

Now let's put real prices on real tools. Start with the flat-seat general assistants, the text engines.

ChatGPT, from OpenAI. Free is zero. Go is eight dollars a month. Plus is twenty. Pro comes in at a hundred and at two hundred dollars a month. Business runs about twenty-five to thirty dollars per user per month with a two-seat minimum, and that's the plan they used to call "Team," renamed Business in August of last year. Enterprise is custom. What do you get where? Plus includes the flagship model, GPT-5.5, which has been the default since April 2026, plus Deep Research at around ten runs a month, image generation, and agent mode. Pro at two hundred dollars gives you roughly twenty times the Plus limits, two hundred fifty Deep Research runs a month, and about a one-million-token context window. Enterprise has no list price. Third-party analysis says negotiated 2026 deals land around fifty to sixty dollars per user per month at a hundred fifty-plus seats, dropping toward forty dollars at five thousand seats. A thousand-seat deal comes out around six hundred thousand dollars a year. Approximate, but it gives you the shape.

Claude, from Anthropic. Trackers of the pricing page show Pro at about twenty dollars a month. Max comes in two flavors: the five-times tier at a hundred dollars, and the twenty-times tier at two hundred. For teams, Team Standard is twenty-five dollars per seat monthly, or twenty on an annual commit. Team Premium is a hundred twenty-five dollars per seat monthly, or a hundred annual. Five-seat minimum on the team plans.

Google's Gemini is a different story now. Google folded Gemini into Workspace and killed the standalone add-ons. To absorb Gemini, they raised the base Workspace prices seventeen to twenty-two percent. So on the business side, per user per month annual, you're looking at roughly seven dollars for Starter, fourteen for Standard, twenty-two for Plus. On the consumer side, Google AI Plus is about five dollars, AI Pro is twenty, and AI Ultra is now a hundred or two hundred dollars. That's after Google cut Ultra from two hundred fifty down to two hundred at their I/O event this year and added a hundred-dollar tier underneath it.

Then the dedicated marketing copy tools. Jasper is fifty-nine dollars per seat per month annual, sixty-nine monthly, and its Business tier is quote-only, with third-party estimates ranging from around nine hundred dollars a month up to six thousand and beyond. Copy.ai has a free tier at two thousand words a month, a Pro plan at forty-nine dollars a month, or thirty-six annual, covering up to five seats, and then Team and Enterprise are quote-only. Writer starts at twenty-nine dollars per user per month annual, thirty-nine monthly, and its Enterprise is negotiated. The contract data floating around: ten to fifty thousand a year for small deals, seventy-five to two hundred fifty thousand mid-market, and five hundred thousand and up for large.

Now flip to the credit-metered side, because this is where marketers get surprised.

Canva first, because it's the cautionary tale. The flat tool that started metering its AI. Canva Pro is fifteen dollars a month, or a hundred twenty a year, and it includes five hundred AI credits a month across their Magic Studio features, Dream Lab, Magic Write, Magic Eraser, Magic Expand. Here's the catch. Those credits reset monthly. They do not roll over. And on Pro, you can't buy more à la carte. The old flat Teams plan is closed to new signups. Heavy users have to buy a separate AI Pass add-on, which gives forty times the Pro credits. So a tool you thought was flat now has a meter running inside it. That's the pattern to watch for everywhere.

Midjourney meters by fast GPU-hours. Basic is ten dollars a month for three-point-three fast hours. Standard is thirty for fifteen hours. Pro is sixty for thirty hours. Mega is a hundred twenty for sixty hours. Twenty percent off if you pay annually. Unused fast hours expire every cycle, no rollover. But every paid tier gets unlimited slow "relax" mode, which matters a lot for how you work, and we'll come back to it. Overage is four dollars per fast hour. And a real gotcha: firms over a million dollars in revenue have to be on Pro or Mega to use it commercially. One more thing to sit with: that ten-dollar Basic plan's fast hours run out in about three hours and twenty minutes of actual generation. Not three hours a day. Three hours, total, for the month.

Adobe Firefly is credit-metered inside Creative Cloud. Standalone, Standard is nine ninety-nine a month for two thousand premium credits, Pro is nineteen ninety-nine for four thousand, and Premium is a hundred ninety-nine ninety-nine for fifty thousand. The nice part: paid plans get unlimited standard image generations. Credits only get burned by premium features, like video and the higher-end models. No rollover. Enterprise and API run roughly two to ten cents an image with about a thousand-dollar-a-month minimum.

Runway, for video, charges credits per generation. Free gives you a hundred twenty-five one-time credits. Standard is twelve dollars per user per month annual for six hundred twenty-five credits a month. Pro is twenty-eight for twenty-two hundred fifty. Max is seventy-six for ninety-five hundred. To make that real: Standard's six hundred twenty-five credits gets you about twenty-five seconds of Gen-4.5 video. A month. So video generation eats credits fast, and you have to plan around it.

ElevenLabs, for voice, meters by character credits. Free is ten thousand characters. Starter is six dollars for thirty thousand. Creator is twenty-two dollars for a hundred twenty-one thousand. Pro is ninety-nine for six hundred thousand. Scale is two hundred ninety-nine for one-point-eight million. Text-to-speech costs one credit per character. Commercial use requires at least the six-dollar Starter plan. And here's the one bright spot in the whole credit world: ElevenLabs rolls your credits over, up to two months, capped at three times your monthly quota. That's the exception. Canva, Midjourney, and Firefly all make you use it or lose it.

Let me go one layer deeper, to the raw programming interface, the API. Most marketers never touch it directly. But you should understand it, because it explains why draft-versus-final economics even exist. Every app tier you just heard is a markup on these raw token rates. These are per-million-token prices.

Anthropic's rates: Opus 4.8 is five dollars in, twenty-five dollars out. Sonnet 5 is three in, fifteen out, with an intro rate of two and ten running through the end of August 2026. Haiku 4.5 is one dollar in, five dollars out. There's a batch option at fifty percent off, and prompt caching that drops cache-hit reads to ten percent of the input cost. Notice the gap. The cheap draft model, Haiku, is five dollars for output. The flagship, Opus, is twenty-five. A five-times spread between the model you draft with and the model you finish with.

OpenAI's rates tell the same story, wider. GPT-5.5 is five dollars in, thirty out. GPT-5.4 is two-fifty and fifteen. GPT-5.4-mini is seventy-five cents and four-fifty. GPT-5.4-nano is twenty cents in, a dollar twenty-five out. Batch is fifty percent off. So from nano's output at a dollar twenty-five up to GPT-5.5's output at thirty dollars, that's about a twenty-five-times span. That span is the entire economic argument for tiering your drafts down to a cheaper model. It's not marketing. It's the raw cost curve.

One footnote for accuracy. Anthropic's docs note that Opus 4.7 and up, and Sonnet 5, use a newer tokenizer that produces about thirty percent more tokens for the same text. So a headline per-token price isn't directly comparable across model generations. Same words, more billed tokens. Just something to know before you do apples-to-apples math.

Now, the heart of the episode. Cost per finished asset.

The reframe is everything. When a tool markets itself, it shows you cost per generation, or cost per API call. Pennies. But that is not what an asset costs you. The real number is the fully-loaded cost of one thing you actually shipped. And that number includes the generations that failed, the human edit time, and the fact-check pass.

Think about the components. There's software licensing, fifty to five hundred dollars a month per tool. There's the labor of prompting and running the AI, call it a quarter to a half of a full-time person. And there's human review. Now watch what happens when we do the arithmetic.

Here's the worked example I want you to remember. Say you've got a hundred-dollar-a-month subscription, and it produces twenty articles a month. Each article needs ninety minutes of editing, and your editor's loaded rate is fifty dollars an hour. So the math is: a hundred dollars for the tool, plus one-and-a-half hours times fifty dollars times twenty articles. That's a hundred dollars plus fifteen hundred dollars. Sixteen hundred dollars a month. Divided by twenty articles, that's eighty-five dollars per article. Not the five dollars the tool's marketing implied. The subscription was six percent of the true cost. The human edit time was about ninety-four percent.

Let that land. Ninety-four percent. The tool is a rounding error. The human hours are the cost.

Now run the same logic on the credit side, because the shape is identical. Say you're making social graphics in a design tool. You've got five hundred credits a month, and one batch of on-brand images eats, say, twenty credits a try. On paper that's twenty-five batches. Feels like plenty. But you don't nail the look on the first try. You run it four times to get one you'll actually use. So your real yield is closer to six keeper batches, not twenty-five. And each of those keepers still needs a human to drop in the headline, check the logo placement, and make sure the hands have five fingers. The credits were never the expensive part. Those four throwaway tries cost you pennies. The person fixing and approving the one you keep costs you the afternoon. Same story as the articles. Different meter, same blind spot.

Some rules of thumb to hang onto. Quality AI content still needs twenty to forty percent of the traditional writing time, just for review. Budget twenty to forty dollars an article for that quality-control pass. And businesses that assume AI output is publish-ready always discover the same thing: five to ten hours a week of review appears out of nowhere and stays. It's a permanent hidden cost, not a startup cost. On the video side, AI-assisted or edited output runs roughly a hundred twenty to six hundred dollars per qualified finished asset.

So here's the copyable workflow. This is the thing to screenshot. How you actually build the cost-per-finished-asset number, in six steps.

One. Count how many generations it takes to get one usable draft, and include the failures, not just the winner. Two. Multiply your raw generation cost by that number of attempts. Three. Add the human edit minutes times the loaded hourly rate. Four. Add the fact-check, brand, and legal pass time. Five. Divide the tool subscription across the whole month's asset count. Six. Sum it all up. That's your cost per finished asset.

And the punchline of that workflow: steps three and four, the human edit and the fact-check, usually dwarf steps one, two, and five combined. The generations and the subscription are the small numbers. The people are the big ones.

If that sounds familiar, it should. Remember the human-in-the-loop editing pass we spent a whole episode on, and the fact-check discipline from the AI-slop episode? Those aren't just quality practices. They are literally line items three and four in this math. The work we told you never to skip is the work that dominates your budget. Which means anything that reduces it is a direct cost lever.

That's the bridge to drafts versus finals.

The engineering world has a concept called model routing. Send the boring, commodity work to the cheapest model that can handle it. Reserve the expensive frontier model only for the pass that actually ships. Brainstorming, first-draft summaries, simple classification, that all goes cheap. The final polished version goes to the good model. Requesty reported that a seventy-twenty-ten split, seventy percent of work on the cheap model, twenty on the mid-tier, ten on the frontier, cuts average cost sixty to eighty percent with no visible quality drop. Teams report bills falling forty to eighty-five percent.

For you, with no code, here's how that translates. Use the free or cheap tier, the smaller model, or fewer credits for all your exploration. The outlines. The ten headline variants. The mood-board images. The rough scratch voiceover. Then switch to the flagship, or the high-res render, or the final credits, only for the one version that ships.

Make it concrete for images. Iterate your composition in Midjourney's relax or slow mode, which is unlimited, or in Firefly's unlimited standard generations. Then spend your precious fast hours or premium credits only on the final upscale. For video, storyboard with short, low-credit generations before you commit to the full render. Remember, six hundred twenty-five Runway credits is only about twenty-five seconds. You don't want to discover the composition is wrong after you've rendered the expensive version.

And the anti-pattern to avoid: regenerating the final-quality asset over and over to fix a small issue you could've fixed with a two-minute manual edit. That's the regeneration spiral, and it burns credits and time at the same time. It's the paid cousin of the AI-slop trap, where you keep hitting generate hoping the machine fixes what a human hand could fix in seconds.

Now, estimating and capping spend, so you're not surprised.

For teams, do the seat math. Headcount times who actually needs it. The dominant waste is idle seats, always. A five-person team on Claude Team Premium annual is six thousand dollars a year. The real question isn't whether to buy it. It's whether all five need Premium, or whether two Premium plus three Standard covers the actual work. That one question can cut the bill meaningfully.

And here's the same move for a team of one. If you're a freelancer or a solo founder, you don't need a team plan at all. You need to ask which single flagship seat, at around twenty dollars, does eighty percent of your work, and then whether any credit-based tool earns a spot beside it. Most people answer that backwards. They collect subscriptions the way you collect browser tabs, one for every task that came up once, and they never close them. So the seat math isn't only a team exercise. A team of one still pays for idle tools, and the idle tool doesn't care that it's the only one you forgot about.

For credits, convert your output plan into credits directly. Say it out loud: "Eight videos a month. Each final render is about X Runway credits, plus Y test generations. That's tier Z." Anchor it with the real rates. Runway, six twenty-five credits is twenty-five seconds. ElevenLabs, one credit per character. Canva, five hundred credits a month. Do that math before you subscribe, not after.

For the technical listener, or your developer, if you touch the API, there are usage caps. OpenAI offers organization monthly caps, per-project budgets, and alert thresholds. But here's a critical 2026 caveat you have to know. OpenAI's monthly budget threshold is a soft cap now. It alerts you, but it no longer hard-stops your requests, and it keeps billing. The only native hard stop is prepaid credits with auto-recharge turned off. So if you think you set a limit, double-check whether it actually stops the bill or just sends you a sad email after the money's gone.

The planning question that ties it together: how many assets a month, at what quality, do I actually need? Price the tier to that honest number. Not to the aspirational maximum you imagine you'll produce.

Let's talk about the hidden costs, because this is the meat.

The biggest one you already know. Human edit and review time. Often ninety percent-plus of your cost per finished asset. If you take one thing from this episode, it's that you're not really buying software, you're buying a thing that generates work for a human. Second, tool sprawl. About one-third of global software spend is wasted on unused seats, duplicate tools, and auto-renewals nobody canceled. Businesses waste up to thirty percent of their software budget on unused licenses, and consolidation can cut the footprint up to forty percent. The marketer version of this is paying for ChatGPT Plus, and Claude Pro, and Jasper, and Copy.ai, and Canva, and Midjourney, all at once, when two tools cover ninety percent of what you do.

Third, regeneration waste. Credits and tokens spent on outputs you never used. Fourth, credit expiry. Canva, Midjourney, Firefly, all forfeit unused capacity every month. ElevenLabs is the exception. And this punishes bursty, seasonal work especially hard, because you pay for a monthly average but you actually work in spikes. Fifth, per-seat costs that don't scale. Sixty-nine-dollar Jasper is fine at two seats and painful at twenty. Sixth, the enterprise pricing cliff. You jump from a self-serve twenty-or-thirty-dollar seat straight to a quote-only, sales-gated Enterprise tier. ChatGPT around forty to sixty-plus a seat with a hundred-fifty-seat minimum. Writer, ten thousand to five hundred thousand-plus a year. Jasper Business, nine hundred and up a month. And remember, opaque pricing means the quote is a negotiating position, not a rate card. Seventh, overage and consumption surprises. Midjourney's four dollars a fast hour. Seventy-eight percent of IT leaders ate an unexpected AI charge. And eighth, commercial-use floors. Midjourney requires Pro or Mega for firms over a million in revenue. ElevenLabs requires the paid Starter plan for any commercial use at all. Which means a "the free tier works fine" plan can actually be legally unusable for your business.

That brings us to build versus buy, and consolidation.

Here's the honest take. A general-assistant subscription, ChatGPT or Claude or Gemini, covers you completely when the job is text. Drafting, ideation, editing, repurposing, building briefs. One twenty-dollar seat replaces a surprising number of specialized writing tools, because Jasper and Copy.ai are largely a prompt-and-template layer sitting on top of the same underlying models you're already paying for. That's worth saying plainly. You may be paying twice for the same engine.

A specialized credit-based tool earns its slot when it does something the general assistant genuinely can't. Image quality, that's Midjourney. Video, that's Runway or Firefly video. Production-grade voice, ElevenLabs. Brand governance and compliance at scale, Writer. The rule is simple: buy the specialist when the specific capability is what you're paying for, not just the convenience.

And the consolidation heuristic: reduce the number of tools before you optimize any single plan. Audit your overlap quarterly. There's also a "build" option for a no-code marketer, which is really the API and automation layer. That's usage-based, pennies per asset at scale. Versus "buy," the packaged app on a flat seat. The API is cheaper per unit but it needs someone technical to wire it up. The app is the markup you pay precisely so you don't need that person. Neither is wrong. Just know which one you're choosing and why.

Let me make the consolidation audit concrete, because "audit your overlap" sounds like homework nobody actually does. Once a quarter, list every AI tool you pay for, and write one sentence next to each. What job does this, and only this, do for me? If two tools answer with the same sentence, one of them goes. If a tool's sentence is just "it writes text," and you already pay for a general assistant, that tool is probably a template layer you can drop without losing anything. The ones that survive the audit are the ones with a sentence no other line item can write. Real image quality. Real video. Real voice. Real brand governance at scale. That's the whole method. One sentence per tool, kill the duplicates, keep the specialists, and do it before you spend an afternoon optimizing a plan you shouldn't be paying for at all.

Let me leave you with the pitfalls, the short list of ways this goes sideways.

Credit anxiety that pushes you to ship a bad first draft. That's the single most likely failure for a listener, under-iterating to save pennies while paying dollars in edit time. Regeneration spirals on final-quality assets. Idle-seat drift, where the team shrinks but the seats don't. Treating "unlimited" as free, when it's only free of marginal token cost, never free of edit time. Trusting a soft cap, like OpenAI's, that won't actually stop the bill. And forgetting about expiry.

So the rules of thumb, if you write nothing else down. Cheap tier for drafts, expensive tier for finals. Count edit hours, not generations. Consolidate before you optimize. And set a hard cap, prepaid with auto-recharge off, not a soft one that just watches the money leave.

The reason this is the last tutorial in Act One is that it's the one that makes everything before it real. The brand voice work, the content brief, the editing pass, the image workflow, they all show up here as either a cost or a savings. A good brief means fewer failed generations, which is fewer attempts in step two of that math. Strong brand voice means less rework in step three. Catching slop early means you don't pay for it twice. The skills weren't just about quality. They were always, quietly, about money. So the next time a tool advertises a price, run it through the six steps before you believe it. The sticker is the smallest number on the page.