Frontier AI capability is collapsing in price every quarter, which cuts your costs and your prices at the same rate; this is how to tell which side of that line your business is actually standing on.
I sell help with intelligence for a living. For three years that felt like standing on rock. This month it started to feel like standing on ice.
On July 9, Meta shipped a frontier agentic model called Muse Spark 1.1 and opened a paid developer API for the first time in the company’s history, on the same day. The price: $1.25 per million input tokens, $4.25 per million output. DeepSeek will run your output tokens for 87 cents a million. And when OpenAI put its newest flagship on stage, the headline pitch was not that it is smarter. It was that it is more efficient.
Read those three facts next to each other and you see the shape of the year. Being smartest stopped being defensible. So everybody pivoted to being cheapest.
Here is the direct answer to the question in the headline, before you scroll another inch. When AI becomes a commodity, two things happen to you at once and they are the same event. Your input costs fall further and faster than you planned, which is a genuine gift. And if what you sell is access to intelligence, your pricing power falls at exactly the same rate, which is not. The gift and the threat arrive in the same box. Which one you actually open depends on one thing: whether your customer is buying the capability from you, or buying your judgment about where to point it.
That is the whole thesis. The value is moving out of having the capability and into knowing which problem to aim it at. Everything below is me working out what that means for you, and, if I am honest, for me.
Key Takeaways
- Frontier AI capability is no longer a durable advantage, and the labs know it; price per outcome is now the battlefield they have all chosen.
- If your business resells access to intelligence, your pricing power is falling at the same rate as your input costs, and the two are mathematically linked.
- Roughly a third of agencies have already been asked for an “AI discount,” and nearly half expect the ask is coming.
- The defensible layer is problem selection, accountability, and taste, because none of those get cheaper when tokens do.
- Falling input costs are a real windfall, but only for businesses that sell an outcome rather than an interface.
The Problem
I built a prompt tool once and watched it die anyway. Good product. Real users at first. Then silence. I have written about that before, so I will not relitigate it here, except for the part that is relevant now: the reason it died is the reason this article exists. I had built a thin layer of convenience on top of somebody else’s capability. When the capability got better and cheaper, my layer got thinner. I did not get outcompeted. I got absorbed.
That is the problem, and it is bigger than a side project. A very large number of businesses built in the last three years are wrappers. Not in the insulting sense. In the structural sense. You found a capability, you understood it before your market did, and you charged for the gap between what the model could do and what your customer knew the model could do. That gap was your margin.
The gap is closing. Not because you got worse. Because the market got smarter and the tools got cheaper, at the same time, on purpose.
And I want to be careful here, because it would be very easy to write this article as though it is about other people. It is not. I run an AI coaching and mentorship business. My entire proposition, stated plainly, is that I know things about AI that entrepreneurs do not, and I will teach them. Every month that AI gets easier, cheaper, and more obvious, the raw information part of that proposition is worth less. If you think I have not sat with that, you have not been paying attention to me.
So this is not a warning I am delivering from safety. This is a thing I am inside of.
Here is the reframe, and it took me a while to get to. The commoditization of capability does not destroy value. It relocates it. Every input that goes to zero makes something adjacent to it more valuable, not less. The question is not whether you survive the price collapse. The question is whether you are standing on the part of the stack that gets cheap, or the part that gets scarce.
Most people cannot answer that about their own business. That is the actual problem.
The Evidence
I do not want you taking my word for the shape of this. Look at what the companies with the most information are doing with their money.
Meta gave away its price position on purpose. Muse Spark 1.1 launched July 9, 2026 with a 1M-token context window, top-of-table scores on tool-use benchmarks like MCP Atlas, and API pricing of $1.25 in / $4.25 out per million tokens. That is roughly a quarter of what comparable frontier models charge. Cline’s CEO Saoud Rizwan said the quiet part in Meta’s own launch post: “strong tool use at a price point that makes it viable to run real coding workloads at scale. That combination is rare.” Meta did not price low because it had to. Meta priced low because capability alone no longer buys you a premium.
DeepSeek is winning on price and publishing the receipts. DeepSeek’s own pricing page lists V4-Flash at $0.14 per million input tokens and $0.28 output, and V4-Pro at $0.435 input and $0.87 output, both with 1M context. Compare that to OpenAI’s GPT-5.6 tiers: Sol at $5 / $30, Terra at $2.50 / $15, Luna at $1 / $6. Sol’s output tokens cost roughly thirty-four times DeepSeek V4-Pro’s. That gap is not a quality gap of thirty-four times. Nobody claims it is.
The flagship pitch changed from intelligence to efficiency. OpenAI positioned GPT-5.6 Sol at the same rate card as GPT-5.5 and led with token efficiency gains rather than a raw capability leap. When the smartest company in the room stops selling smart and starts selling cheap, that is not modesty. That is a read on the market.
Meta wants to sell the shovels, not just the gold. On July 1, Bloomberg reported that Meta is building a cloud business, internally called Meta Compute, to sell its excess AI capacity against AWS, Google Cloud, and Azure. TechCrunch’s read is the sentence I keep rereading: it is “a signal that the winners of the AI race may not be the ones providing the best models and services, but rather the ones who own the data centers.” Meta has committed $182.9 billion to AI infrastructure. When a company that big decides its models are a commodity and its power bill is the asset, believe it.
The price collapse has a track record, not a forecast. Stanford’s AI Index found that querying a model at GPT-3.5 level fell from $20 per million tokens in November 2022 to $0.07 by October 2024. That is a more than 280-fold drop in about eighteen months. Not a projection. History.
And your clients have already done the math. A Productive survey of 180-plus agencies found that around a third have already been asked for an “AI discount,” and nearly half expect the question is coming. Coverage of the same report puts the share who actually cut prices at 13 percent, which tells you something hopeful: the ask is universal, the capitulation is not. One senior account manager in that survey put it exactly right. “Clients expect to pay less because we might use AI, but the price cut they expect doesn’t reflect the reduction in human resources required to complete tasks to a high-quality standard.”
Six data points, one story. The people closest to the technology have concluded that the technology is not the product.
The Solution: The Substitution Test
None of this is new economics. It is just new to us.
In 2002, Joel Spolsky named a pattern that has run through tech for forty years: commoditize your complement. Demand for your product goes up when the price of the thing next to it goes down. So smart companies work relentlessly to drive the price of their complements to zero. IBM commoditized PC add-ins to sell PCs. Microsoft commoditized PCs to sell DOS.
Now look at the last two weeks with that lens. Meta is commoditizing frontier intelligence so it can sell compute. DeepSeek is commoditizing frontier intelligence to buy distribution and mindshare. Every lab is racing to make the thing you resell free, because the thing they actually monetize sits underneath it or beside it.
Here is the uncomfortable part. If intelligence is somebody’s complement, and you are selling intelligence, then you are the complement. You are the thing being commoditized. On purpose. By people with $182.9 billion.
So the move is to figure out what you own that is not somebody else’s complement. I use one question for this, and it is blunt on purpose.
The Substitution Test: For every line item you charge for, ask what it would cost your client to get an 80 percent version of that exact thing, this afternoon, from a $0.14-per-million-tokens model and a competent VA.
Not a 100 percent version. An 80 percent version. Because 80 percent at a hundredth of the price wins most purchase decisions, and pretending otherwise is how businesses die politely.
Run every line through it. Some will come back “impossible,” some “expensive,” and some “twenty bucks.” The twenty-dollar ones are not your business anymore. They are your cost of goods sold. Stop pricing them like assets.
What survives the test is always the same short list, and it is worth noticing that none of it is about the model.
Problem selection. The model will answer any question you ask it, brilliantly, including the wrong one. Knowing which question is worth asking is not a capability, it is a judgment, and judgment does not have a price-per-token curve.
Accountability. Nobody can sue a token. When you attach your name and your money to an outcome, you are selling risk transfer, and risk transfer gets more valuable as the tools get more powerful, not less.
Taste and context. Yonah van Andel, who runs a Dutch agency called On a Daily Basis, said it better than I can in that Productive report: “with AI, anyone can generate a brand strategy in seconds. What will really be rewarded is the point of view behind the tool.”
My own proof point is unglamorous. The parts of my business built on “here is information you do not have” have gotten steadily harder to sell, exactly as this analysis predicts. The parts built on “here is the specific problem in your specific business and here is where to aim this” have not. Same me. Same tools. Wildly different durability. I did not plan that. I noticed it.
Practical Steps
Run the Substitution Test on every line item this week. Write out everything you charge for, one row each. Next to each, write what an 80 percent version costs your client from a cheap model plus an hour of somebody’s time. Be honest to the point of discomfort. The rows that come back cheap are not negotiable, they are already gone.
Reprice the survivors as the whole product. Whatever survived the test is what you actually sell. Everything else is packaging. Most people have this exactly inverted, pricing the deliverable and giving away the judgment for free as a “consultation.” Flip it. The judgment is the product; the deliverable is the receipt.
Take the input windfall, and do not pass it through. Your token costs are dropping, and that money is real. Do not reflexively hand it to clients as a discount because they asked. The Productive data says nearly everyone gets asked and only about one in eight actually cuts. Reinvest yours in depth, speed, or guarantees. Those are things a discount cannot buy back.
Move at least one offer to an outcome. Not hours, not deliverables, not seats. One offer, priced against a result your client can feel in their bank account. Start small enough that you can survive being wrong twice. Outcome pricing is the only pricing that is immune to your inputs getting cheaper, because it was never indexed to them.
Rebuild your pitch around the aim, not the ammunition. If your sales conversation includes the phrase “we use AI to,” you are describing ammunition. Everybody has ammunition now; Meta is handing it out with $20 in free credits. Describe the aim instead: which problem, why that one, and what happens when it is solved.
Answer the discount ask before it arrives. It is coming for you; the data says it comes for everybody. Have a real answer ready, and make it a question: “What would it cost you to be confidently wrong about this for six months?” That reframes the purchase from production cost to decision risk, which is what they are actually buying.
Audit yourself again in ninety days. Inference prices have been falling roughly an order of magnitude a year at a fixed quality bar. That means the Substitution Test has a shelf life measured in months. What passes today fails by Q4. Put it on the calendar; do not wait to feel the squeeze.
Frequently Asked Questions
Is AI actually getting cheaper, or is that just marketing?
It is real and it is documented. Stanford’s AI Index found the cost of querying a GPT-3.5-level model fell from $20 per million tokens in November 2022 to $0.07 by October 2024, more than a 280-fold drop in roughly eighteen months. Meta’s July 2026 frontier model launched at about a quarter of comparable frontier pricing. The trend is measured, not projected.
Should I switch to DeepSeek to save money?
Maybe, but that is a smaller question than it looks. DeepSeek V4-Pro lists at $0.435 input and $0.87 output per million tokens versus GPT-5.6 Sol at $5 and $30, so the savings are real. But cutting your own costs does not protect your prices. If your business is exposed to commoditization, a cheaper model makes you profitable for slightly longer, not safer.
My clients are asking for an AI discount. Do I give it?
Around a third of agencies have already faced that ask and nearly half expect it, yet only about 13 percent have actually cut prices. That gap is the answer. The efficiency gain is yours; you took the risk of learning the tools. Reinvest it in better outcomes rather than surrendering it, and reframe the conversation from production cost to decision quality.
Does this mean AI consulting and coaching businesses are doomed?
No, but the information-arbitrage version of them is. Selling “I know things about AI that you don’t” has a shrinking half-life because the tools keep getting more obvious. Selling “I know which problem in your business is worth pointing this at, and I will stand behind the result” does not commoditize, because judgment has no price-per-token curve.
What is actually safe from AI commoditization?
Three things, consistently: problem selection, accountability, and taste. A model will answer any question you ask, including the wrong one. It cannot choose the question, cannot be liable for the answer, and cannot hold your client’s context in a way they trust. Everything else in your offer is on the clock.
The Close
I told you at the top that I sell help with intelligence and that it started feeling like standing on ice. I want to finish that thought honestly, because I think the honesty is the useful part.
The ice is real. The part of my business that was ever about knowing things first is melting, and it is melting because a trillion dollars of capital decided that knowing things should be free. I cannot out-spend that. Neither can you. Nobody reading this is going to win a price war against Meta’s spare capacity.
But here is what I keep coming back to, and it is the thing I actually believe.
Every single one of those companies just told us, with their pricing pages, that the intelligence is not the valuable part. They would not be giving it away otherwise. They are commoditizing the exact thing everybody spent three years being afraid of. And the thing they cannot commoditize, the thing that is not on any of their rate cards, is a human being who knows their business well enough to know which problem is worth solving first.
That was always the job. We just got distracted for three years because the tools were shiny and the arbitrage was easy.
So take the windfall. Your costs are falling further than you planned and that money is yours. But stop selling the ammunition. There is an infinite supply of ammunition now and it costs fourteen cents a million.
Sell the aim. Nobody is giving that away.
If this is the kind of thinking you want more of, come find me. I work through this stuff in public, out loud, including the parts where I am the one exposed. Follow along and let’s figure out where to aim.
About the author: Jonathan Mast is the founder of White Beard Strategies, where he provides AI coaching and mentorship to entrepreneurs who would rather use this technology than be replaced by it. He is a speaker, a builder, and the creator of the Perfect Prompt Framework. He writes about AI and business with a bias toward what is actually true over what is comfortable, and he has the failed products to prove it.
