DeepSeek V4 Flash matches frontier models like Claude Opus 4.8 on coding and prices its output in cents per million tokens against roughly $25, call it a 99% discount. Qwen3.8-Max is doing the same thing open-weight, and OpenAI has cut GPT-5.6 prices by around 80%. Every “AI makes marketing cheaper” deck is about to cite these numbers. Nearly all of them are making the same category error.
The falling number is the cost of intelligence, the cost of producing marketing. The number that decides whether you have a business is the cost of demand, the cost of getting a customer to choose you. Cheaper models do nothing to move the second one. In fact, AI search is pushing it up. The model got 99% cheaper. Your CAC did not get the memo.
The number everyone’s about to misread
The price war is real and it is dramatic. A model ranked at or near the top of a crowdsourced coding board is charging a rounding error against what the frontier labs charged a year ago. An open-weight competitor is matching closed-model quality and handing the weights away. A major lab cut its mid-tier prices by roughly 80% in a single move. Commentators have reached for the obvious metaphor: intelligence is commoditizing, priced like electricity or gasoline.
That framing is correct, and it is about to be badly applied. Because “the cost of AI is collapsing” is going to get translated, in a hundred strategy decks, into “the cost of marketing is collapsing.” Those are not the same sentence. They are not even about the same cost.
The category error
There are two different costs, and the price war only touches one of them.
The first is the cost of intelligence: what it takes to produce the asset. The draft, the variant, the image, the brief, the line of code. This is what DeepSeek and Qwen and the GPT-5.6 cut are crushing. Producing a competent marketing artifact has never been cheaper, and it is getting cheaper by the week.
The second is the cost of demand: what it takes to get a human being to notice you, trust you, and choose you over the alternative. That cost has nothing to do with token pricing. A customer does not become cheaper to acquire because the ad that reached them was drafted by a model that costs a fraction of last year’s. Collapse these two costs into one line item and every conclusion that follows is wrong.
What actually got cheaper
Give the productivity story its due, precisely, so the rest of the argument is not mistaken for hype-skepticism. Real things got cheaper, and you should take the savings.
First drafts are nearly free. Variant production, ten versions of a headline, twenty cuts of a creative, is trivial now. Briefs, outlines, first-pass creative, research synthesis, and a great deal of code all dropped to a fraction of their former cost in time and money. This is genuine, and a team that refuses to use it is leaving efficiency on the table for no reason. The point is not that the cheaper model is worthless. The point is exactly what it is worth: the production half of the work, and only that half.
What didn’t, and is getting more expensive
Now the other half. Attention, trust, and incremental demand did not get cheaper, and the same force making production cheap is making them more expensive.
When the cost of producing content falls toward zero, the market floods with adequate content. Everyone’s competitor can now generate the same competent blog post, the same passable ad, the same serviceable landing page, for pennies. The moment everyone can produce it, producing it stops being an advantage. The scarce input is no longer the content, it is the distribution to get it seen and the distinctiveness to make it chosen. Commoditized supply always does this: it moves the value to whatever is still scarce, and what is still scarce is human attention and trust.
AI search is pushing demand cost the wrong way
This is where it connects to what is already happening in search. AI answers compress the cheap-acquisition funnel, the informational query that used to send a free click to your page now gets resolved inside the answer, and paid gets forced back upstream to buy the visibility that organic used to earn. So the cost of intelligence is falling while the cost of being discovered is rising, at the same time, driven by the same wave of AI adoption.
Read those two trends together and the “AI makes marketing cheaper” thesis inverts. Yes, the asset got cheaper to make. But the environment you have to place that asset into to win a customer got more crowded and more expensive to reach. Cheaper to produce, harder to be chosen. That is not a cost reduction. That is a cost shift, toward the part of the job AI cannot do for you.
The measurement trap this sets up
There is a specific way this goes wrong on a dashboard, and it is worth naming so you can catch it. Cheaper content leads to more content, which leads to more activity, which leads to reports that look busier than ever. More posts shipped, more variants tested, more campaigns live. The dashboard fills up.
And incremental demand can be dead flat underneath all of it. Volume is not lift. This is the same failure mode as platform ROAS grading its own homework: the metric that is easy to move gets moved, and everyone mistakes motion for progress. A holdout does not care how cheap the token was. It measures whether the activity caused demand that would not have existed otherwise, and cheap production has a way of inflating everything except that number. If you let output volume become the KPI, the price war will make your reports look spectacular while your CAC quietly climbs.
So what do you actually do with a 99%-cheaper model?
Take the savings, and reinvest them in the half that got more expensive. The teams that win the next couple of years will treat cheap intelligence as a reason to spend more on the things it cannot commoditize: distinctiveness, so your adequate-content competitors do not blur into you; distribution, so the asset is actually seen by the right person at the right moment; and measurement that proves lift, so you know which of your now-abundant activity is doing real work.
Cheaper production is only an edge if you spend the surplus where demand is actually won, and demand is won at the level of intent, on the right channel, measured on one honest scoreboard rather than a busier one. Pocket the token savings and pour them into the expensive half. That is the whole play.
Close
Commoditized intelligence is good news, for your cost line, not your moat. When the model gets 99% cheaper, so does your competitor’s model, which means the cheapness is not an advantage anyone keeps. The advantage moves to whatever stays scarce: attention, trust, and provable incremental demand. The teams that understand this will use cheap models as a reason to invest more in the parts AI cannot touch. The teams that misread it will generate more content than ever and wonder why nothing moved.
So if producing marketing just got 99% cheaper, here is the real question: what is the one line item you would move that budget into, and why that one?