TL;DR: Yes, but the question as usually asked is the wrong one. Whether a product built on someone else’s model is a business is settled almost entirely by two numbers that have nothing to do with how thin the product is: which model tier it calls, and how heavy its heaviest users are. Priced from published August 2026 rates, a realistic job on the cheapest capable tier costs $0.52 per thousand jobs, and the same job on the top tier costs $40.00 per thousand, a 77-fold spread within the same market. On a $20 monthly plan that is the difference between 92% gross margin and minus 140%. Two further findings cut against the conventional wisdom. First, prices are not uniformly collapsing: one vendor’s flagship tier fell 67% while its cheapest tier rose 25%, and a tokenizer change documented by that vendor means roughly 30% more tokens for the same text, which cuts a headline 33% discount down to 13% on the workhorse tier. Second, the demand side is the opposite of what wrapper-shaming implies: in MIT’s 2025 enterprise study, tools bought from outside vendors reached deployment about 67% of the time against about 33% for internal builds. Buyers are not trying to avoid your wrapper. They are trying to avoid building it themselves.
The insult arrived before the analysis did. Call a product a wrapper and the conversation is over, which is convenient, because the word has never meant anything precise. Every software business sits on infrastructure it does not own. The question was never whether you wrap something. It is whether the thing you add is worth more than it costs you, and that is arithmetic.
So let us do the arithmetic, with published prices, and then look at what buyers actually did.
What the model tier costs you, in the only unit that matters
Token pricing is quoted per million tokens, which is a unit nobody’s business runs in. Businesses run in jobs: one user action that triggers one meaningful model call.
Define a job as 4,000 input tokens and 800 output tokens. That is a system prompt, a document or conversation slice, and a substantive answer. It is neither the cheapest nor the most expensive shape of real usage, and it lets us compare tiers on the same basis.
The first table is verified against the vendors’ own published pricing pages as of August 2026. The second and third columns of the second table are derived by us.
| Model | Input, per million tokens | Output, per million tokens |
|---|---|---|
| GPT-5 nano | $0.05 | $0.40 |
| GPT-4.1 nano | $0.10 | $0.40 |
| GPT-4o mini | $0.15 | $0.60 |
| GPT-5 mini | $0.25 | $2.00 |
| Claude Haiku 4.5 | $1.00 | $5.00 |
| GPT-5 | $1.25 | $10.00 |
| Claude Sonnet 5 | $2.00 | $10.00 |
| GPT-4.1 | $2.00 | $8.00 |
| GPT-4o | $2.50 | $10.00 |
| Claude Opus 5 | $5.00 | $25.00 |
| GPT-5.5 | $5.00 | $30.00 |
Now the same prices expressed as what a product owner actually pays. Cost per job is derived: input tokens times input rate plus output tokens times output rate, at 4,000 and 800 tokens.
| Tier | Cost per job | Cost per 1,000 jobs | Monthly cost, user doing 300 jobs | Monthly cost, user doing 3,000 jobs |
|---|---|---|---|---|
| GPT-5 nano | $0.00052 | $0.52 | $0.16 | $1.56 |
| GPT-5 mini | $0.00260 | $2.60 | $0.78 | $7.80 |
| Claude Haiku 4.5 | $0.00800 | $8.00 | $2.40 | $24.00 |
| GPT-5 | $0.01300 | $13.00 | $3.90 | $39.00 |
| Claude Sonnet 5 | $0.01600 | $16.00 | $4.80 | $48.00 |
| Claude Opus 5 | $0.04000 | $40.00 | $12.00 | $120.00 |
The spread between the cheapest and most expensive row is 77 times. This is a CEOtudent editorial calculation from published rates, and the assumptions are stated so you can substitute your own: change the job shape and every number moves, but the ratio between tiers barely does.
The margin table that decides whether you have a business
Put those costs against the most common consumer software price in the world, a $20 monthly subscription, and the picture stops being abstract. Gross margin here is revenue minus inference cost only, before hosting, payment processing, support and the founder’s time.
| Tier | Gross margin, 300 jobs/month | Gross margin, 3,000 jobs/month |
|---|---|---|
| GPT-5 nano | 99.2% | 92.2% |
| GPT-5 mini | 96.1% | 61.0% |
| Claude Haiku 4.5 | 88.0% | negative 20.0% |
| GPT-5 | 80.5% | negative 95.0% |
| Claude Sonnet 5 | 76.0% | negative 140.0% |
| Claude Opus 5 | 40.0% | negative 500.0% |
Read the right-hand column carefully, because it contains the actual failure mode of the last two years of small AI products. Nothing in it is about being a wrapper. It is about flat-rate pricing meeting a usage distribution with a long tail. A product where the top 5% of users generate ten times the median load will look healthy on average and lose money on the customers who like it most.
Three consequences follow, and they are the whole strategy.
Model selection is a pricing decision, not an engineering one. Moving a workload from Sonnet 5 to GPT-5 mini cuts cost per job by roughly 84%, and if the task is classification, extraction, routing or summarizing to a template, the quality difference at the point of use is frequently invisible to the customer. Founders reach for the stronger tier for reasons of comfort and then discover that the comfort costs 6.2 times the inference bill for the same work.
Flat pricing needs a ceiling. Not a hostile one. A generous included allowance with transparent overage, or a hard cap with an upgrade path, converts the tail from an existential risk into a revenue line.
Caching and batching are margin, not optimization. On the Anthropic side a cache hit is priced at 0.1 times the base input rate and batch processing carries a 50% discount on both input and output. For a product with a large stable system prompt hit repeatedly, that is not a rounding error, it is the difference between two of the rows above.
The price curve is not doing what everyone says it is doing
The standard argument runs: inference costs are collapsing, therefore wrapper margins expand automatically, therefore build now. Stanford HAI’s AI Index supports the premise at the historical level, reporting that the cost of inference for performance at the level of GPT-3.5 fell by a factor of more than 280 between late 2022 and late 2024.
But that is a statement about the falling price of yesterday’s capability, not about the price of the frontier. Compare successive generations within a single vendor’s published price list and the picture is much less tidy. The percentage columns below are derived by us from the published rates.
| Move | Input price change | Output price change | What it shows |
|---|---|---|---|
| Claude Opus 4.1 to Opus 5 | $15.00 to $5.00, down 66.7% | $75.00 to $25.00, down 66.7% | The flagship tier repriced sharply downward |
| Claude Sonnet 4.6 to Sonnet 5 | $3.00 to $2.00, down 33.3% | $15.00 to $10.00, down 33.3% | The workhorse tier fell by a third |
| Claude Haiku 3.5 to Haiku 4.5 | $0.80 to $1.00, up 25.0% | $4.00 to $5.00, up 25.0% | The cheap tier got more expensive, not less |
| GPT-4o to GPT-5 | $2.50 to $1.25, down 50.0% | $10.00 to $10.00, unchanged | Input halved, output flat |
| GPT-4o mini to GPT-5 nano | $0.15 to $0.05, down 66.7% | $0.60 to $0.40, down 33.3% | The floor kept falling |
| GPT-5 to GPT-5.5 | $1.25 to $5.00, up 300.0% | $10.00 to $30.00, up 200.0% | Newer frontier tiers priced above the previous flagship |
Two of those rows point upward. Anyone planning a business on the assumption that their input cost falls every year is planning on an average that contains its own counterexamples.
Then there is a footnote most cost models never open. Anthropic’s pricing documentation states that Claude 4.7 and later models use a newer tokenizer that produces approximately 30% more tokens for the same text. That matters because you are billed per token but you ship characters. Applying the vendor’s own approximate figure to the two moves above gives a materially different discount.
| Move | Headline per-token cut | Adjusted for roughly 30% more tokens | Real cut on the same text |
|---|---|---|---|
| Opus 4.1 to Opus 5, input | 66.7% | $5.00 times 1.3 equals $6.50 effective, against $15.00 | 56.7% |
| Sonnet 4.6 to Sonnet 5, input | 33.3% | $2.00 times 1.3 equals $2.60 effective, against $3.00 | 13.3% |
The honest caveat is that the vendor calls the 30% approximate and workload-dependent, so treat these as an illustration of the direction rather than a precise figure for your workload. But the direction is the point: of the headline 33.3% Sonnet discount, 13.3 points survive contact with the tokenizer and 20 points do not. If your unit economics depend on a headline price cut, measure your own token counts before and after a model migration rather than trusting the press release.
The demand-side finding that contradicts the whole genre
Everything above is supply side. The more interesting evidence is about what buyers did, and it runs the opposite direction to the wrapper sneer.
MIT’s NANDA initiative published The GenAI Divide: State of AI in Business 2025, built from 52 structured interviews, a survey of 153 leaders, and a systematic analysis of over 300 publicly disclosed AI initiatives. Its headline finding travelled widely: about 95% of organizations were getting zero return from their generative AI pilots.
The finding that did not travel is the one that matters if you are deciding what to build. In their sample, external partnerships using learning-capable, customized tools reached deployment about 67% of the time, compared with about 33% for internally built tools. The report is explicit that these are self-reported outcomes that may not account for all confounding variables, and the sample is small enough that this should be treated as a strong signal rather than a measurement. But the direction was consistent across interviewees, and it says something unambiguous: the companies that tried to skip the vendor and build it themselves failed roughly twice as often.
The report is equally specific about what buyers were paying for. Executives wanted systems that learn from feedback, raised by 66% of them, and that retain context, demanded by 63%. The organizations succeeding were those working with vendors solving for learning, memory and workflow adaptation. The ones failing were, in the report’s words, either building generic tools or trying to develop capabilities internally.
Read those two sentences together and the strategic instruction is not subtle. Generic is the thing that fails. Building it in-house is the other thing that fails. A narrow outside product that accumulates context about one workflow is the shape that worked.
There is one more number worth having. The same study found roughly 70% of AI budget allocation going to sales and marketing, while noting better returns in operations and finance. If you are choosing a beachhead, the crowded end of the market is the one everyone is funding and the underserved end is the one that pays back.
The market these buyers sit in is not small, and it is not shrinking. The AI Index 2026 puts organizational AI adoption at 88% of surveyed organizations, with generative AI in use in at least one business function at 70% of them, and reports global corporate AI investment of $581.7 billion in 2025, up 130% year over year. That combination, near-universal adoption alongside a 95% pilot failure rate, is the actual market description: an enormous number of organizations have decided they must do this and have not yet found something that works. That gap is the product opportunity, and it is a demand for something specific rather than for another general assistant.
The Wrapper Durability Score
The useful question is not whether a product is a wrapper but how long it survives contact with three forces: the model absorbing its feature, a competitor cloning it in a weekend, and its own unit economics. The following scorecard is a CEOtudent editorial framework. Score each factor 0, 1 or 2 and total it.
| Factor | 0 points | 1 point | 2 points |
|---|---|---|---|
| Proprietary data loop | The product knows nothing today that it did not know at signup | Accumulates user history that improves output within one account | Accumulates a corpus across accounts that no new entrant can replicate |
| Workflow depth | A text box and a button | Integrates with one system the user already lives in | Sits inside a process, holds state, and is the system of record for a step |
| Absorption distance | The core function is a prompt the model could ship as a default | Adds orchestration, evaluation or guardrails a chat interface lacks | Depends on permissions, integrations or compliance a general assistant will not obtain |
| Unit economics headroom | Frontier tier, flat price, no cap | Mid tier with caching, or usage-linked pricing | Cheap tier sufficient for the task, or pricing that scales with usage |
| Distribution that is not paid | Discovery is entirely ads or launch-day attention | An owned audience or an integration marketplace listing | Embedded where the buying decision happens, or referred by the workflow owner |
| Switching cost | Cancel and lose nothing | Lose saved settings and history | Lose accumulated data, integrations and team process |
Interpretation, and this is our judgment rather than a measured threshold: 9 and above describes a product with a defensible year ahead of it. From 5 to 8 describes a real business that has to keep moving, which is most software. Below 5 describes a feature that is renting time until someone ships it as a default, and the correct move there is usually not to abandon it but to spend the next quarter converting one zero into a two.
The single highest-leverage row is the first. A proprietary data loop is what the MIT buyers were describing when 66% of them asked for systems that learn from feedback, and it is the only factor on the list that a well-funded competitor cannot buy on the day they decide to.
What not to build in 2026
Three shapes have a poor record and a worse outlook.
The single-prompt utility. A product whose entire function is one instruction plus a nicer text box competes directly with a default. It also has no proprietary loop, which means every day of operation leaves it exactly as defensible as day one.
The unpriced agent. Agentic products multiply token consumption by the number of steps in a loop, which turns the right-hand column of our margin table from a tail risk into the base case. If a single user request can trigger fifty model calls, a flat monthly price is not a pricing model, it is a bet on low engagement.
The general assistant for everyone. This is the position the MIT data describes as generic, and it competes with the free consumer products that dominate ad-hoc use. Narrow beats broad here, and it is not close.
For adjacent decisions we have covered elsewhere: the pricing structure question is worked through in our comparison of the one-product business against a revenue stack, the getting-started problem in the cold start problem for digital products, the price-setting method in digital product pricing in the AI era, and the maintenance reality of anything sold as hands-off in the passive income myth. The broader earnings picture is in AI side hustle economics and what the data shows.
So, can you still build one
Yes, under conditions that are now specific enough to check before you write code.
Pick a task the cheapest capable tier can do, because that choice is worth up to 77 times in cost and decides your margin before anything else does. Price so that heavy usage increases revenue rather than only cost. Choose a workflow narrow enough that context accumulates into something a general assistant will not have. Sell to a buyer who has already discovered that building it internally fails about two times out of three. And measure your own token counts across a model migration rather than believing a headline discount.
That is not a wrapper strategy or an anti-wrapper strategy. It is the same discipline that made software businesses work before any of this, applied to a cost structure that is genuinely new, and it is available to a single operator with no funding and a month of evenings.
Frequently asked questions
Is a wrapper business inherently less defensible than traditional software?
Not inherently, but the default version of it is. Traditional software also runs on infrastructure it does not own. The difference is that a thin AI product often has no accumulating asset, so it does not get harder to displace over time. Our scorecard’s first row is the fix, and it is available to any product that starts storing what it learns.
What happens when the model provider ships my feature?
It is a real risk and it is concentrated in one row of the scorecard. Features that are a prompt away from a default get absorbed. Products that depend on permissions, integrations, compliance boundaries or accumulated customer data do not, because a general assistant would have to acquire those relationships one customer at a time.
Should I use the cheapest model for everything?
No. Use the cheapest model that passes your own evaluation on your own task, which for classification, extraction and templated writing is often far below the frontier. Route the small share of requests that genuinely need reasoning to a stronger tier. Mixed routing is why the cost-per-job table matters: it lets you price the routing decision instead of guessing at it.
Is the 95% failure figure from the MIT report a reason not to build?
It is a reason to be careful about which claim you take from it. The 95% describes enterprise pilots failing to produce measurable profit and loss impact, from a sample of 52 interviews and 153 survey responses, which is a small base. The more actionable finding in the same report is the 67% against 33% gap between bought and self-built tools, which is an argument for being the vendor rather than against building anything.
How much can prompt caching actually change the numbers?
On the Anthropic pricing schedule a cache hit costs 0.1 times the base input rate. For our example job, where 4,000 of the 4,800 tokens are input and much of a real system prompt is stable across calls, moving most input tokens to cache reads takes the input side of the cost toward a tenth of its listed value. It does not touch output cost. Whether that flips a row in the margin table depends on how much of your input is genuinely repeated, which is a question about your product, not about the pricing page.
Sources
Stanford Institute for Human-Centered Artificial Intelligence, The 2026 AI Index Report, chapters on Technical Performance and Economy, 2026.
Stanford Institute for Human-Centered Artificial Intelligence, The 2025 AI Index Report, on the decline in inference cost for GPT-3.5-level performance.
MIT NANDA, The GenAI Divide: State of AI in Business 2025, July 2025, including its stated research methodology and sample limitations.
Anthropic, published Claude model pricing documentation, including model rates, prompt caching multipliers, batch processing discounts and the tokenizer note for Claude 4.7 and later models, accessed August 2026.
OpenAI, published API pricing documentation for the GPT-5, GPT-4.1 and GPT-4o model families, accessed August 2026.
This content was compiled with the support of AI following in-depth research, then written and prepared for publication by the CEOtudent editorial team.
This post is also available in:
