Google's published rate card lists Gemini 4 Argon at $2 per million input tokens and $10 per million output tokens, and from December 1 those figures double to $4 and $20. Founders who shipped a product on top of that model inherit the change without a vote. Their customers' subscriptions renew at the same price, their support costs stay flat, and the only line that moves is the one that decides whether the business is worth running. It moves in the wrong direction. That is the shape of the AI software era for small operators, where the cost of goods is set on somebody else's pricing page.
Why Wrapper Margins Collapse Under Usage
According to ICONIQ Growth's 2026 State of AI report, roughly $230,000 of every $1 million in AI product revenue leaves the building as inference cost before an engineer, a salesperson or a marketer gets paid. Put that against the old SaaS standard and the problem becomes obvious. Conventional software ran at 70 to 85 percent gross margin because serving customer number two thousand cost almost nothing more than serving customer one. A 77 percent margin is a decent hardware business. In software it is a warning.
The tools that skipped that discipline are already showing the damage. Thin wrappers, the products that put a chat box in front of a general model and call it a company, are running gross margins of 25 to 35 percent against 70 to 85 percent for conventional software. Forecasts tracking that segment put nine in ten of them out of business by the end of this year. Most of them did not fail at building. They failed at pricing, because the price was set before anyone measured what a request actually costs.
The public market has already repriced the category. The median annual recurring revenue multiple across the SaaS Capital Index hit decade-plus lows in the first quarter of 2026 as investors began treating AI as a threat to software business models rather than a gift to them. Private buyers of small SaaS are just as skittish. A founder walking into an acquisition conversation with 40 percent gross margin and no usage cap is negotiating from the floor, whatever the growth chart says.
The Token Math Behind a $29 Plan
Take a support triage tool that reads an 8,000 token ticket and writes an 800 token summary plus a routing note. That is a modest workload by current standards. It sits well below the 272,000 token repricing threshold that every GPT-6 model hits, and below the 100,000 token cliff on Claude Haiku 5.5, where per-token rates change mid-request. Run 2,000 of those requests a month against a $29 plan and the model tier alone decides whether the account is a 92 percent margin or a monthly donation.
| Model tier | List rate per 1M tokens | Cost per 1,000 requests | COGS on 2,000 requests | Gross margin on a $29 plan |
|---|---|---|---|---|
| GPT-6 Astra, Claude Fable 5.1 | $10 in / $50 out | $120.00 | $240.00 | Negative |
| GPT-6.1 Terra | $1 in / $5 out | $12.00 | $24.00 | 17 percent |
| Gemini 3.8 Flash | $0.75 in / $3.75 out | $9.00 | $18.00 | 38 percent |
| DeepSeek V4.1 Flash | $0.30 in / $1.20 out | $3.36 | $6.72 | 77 percent |
| GPT-6 Luna, Claude Haiku 5.5 | $0.10 in / $0.50 out | $1.20 | $2.40 | 92 percent |
Keep Reading
Should You Quit Chasing Micro-SaaS Ideas?The customer receives the same summary and the same routing decision in every row. Nothing about the product changes as you move down that table except the invoice at the bottom, and the spread between the first row and the last is a factor of 100 on identical work. No pricing page survives a 100x cost variance. That is why routing belongs in the revenue conversation and not just the engineering backlog, and why founders who treat model selection as a taste question end up subsidising their own heaviest users.
The table also ignores payment processing, hosting and support, all of which still come out of the same $29. At standard card rates, roughly $1.14 of that subscription disappears before the first token is billed. Add a database and a logging stack and the realistic ceiling on a $29 seat is around $26 of usable gross profit, which is the number the model choice has to fit inside.
Route, Cache, and Cap Before You Ship
The operators who keep margins above 70 percent tend to run the same short list of controls before they take a single paying account.
- Route by task class. Cheap models handle extraction, classification and first drafts, while the flagship tier is reserved for the slice of requests where a human can audit the quality difference.
- Cache the context. Cached input reads bill at a fraction of the standard input rate across the major providers, and most products resend the same system prompt thousands of times a day.
- Batch anything that isn't interactive. Overnight report generation and weekly digests can run through batch or off-peak windows at lower rates.
- Publish a token ceiling per account and price the overage. Goodwill on usage is how a $29 customer becomes a $240 liability.
- Watch the long-context cliffs. A prompt that crosses a provider's threshold repriced the entire request, not just the excess tokens.
- Log cost per account from day one. If you cannot see which customer is eating the margin, you cannot price your way out of it.
Pricing follows the same logic. Per-seat plans made sense when delivery cost was flat, and a metered component restores the link between what a customer consumes and what they pay. Vendors bundling AI as an upsell are charging premiums of 30 to 110 percent over base plans, which only holds if the metered cost sits comfortably beneath the premium. Outcome pricing goes further and bills against resolved tickets or closed books rather than tokens, shifting variance onto the vendor in exchange for a much higher price point.
The cultural shift matters as much as the mechanics. Founders have already stopped treating growth as the only metric, and the ones who internalised that early are pricing on cost to serve rather than on what the market will bear for one quarter. A capped plan with a clear overage rate closes better than a generous one that threatens the business, because the buyer usually knows their own volume better than you do.
Vertical Depth Is the Moat That Survives
Vertical software is a $157 billion market growing at 18 to 22 percent a year, roughly two to three times the pace of horizontal tools. The reason is not that industry buyers are underserved in some abstract sense. It's that the work inside a dental group, a freight broker or a regional clinic involves codes, integrations and payroll rules that a general model has no reason to know. Vertical depth turns a prompt into a data model, and a data model takes longer to clone than a prompt does.
The moat shows up in unglamorous places. A billing tool that already maps a specific insurer's reimbursement rules has to be rebuilt from scratch by a competitor, not reprompted. Every integration, every exception and every export format that a customer relies on becomes switching cost, and switching cost is what lets you raise price without losing the account. Compare that with a generic writing assistant, where the buyer can leave in an afternoon and the only thing you own is their login email.
Governance has turned into a sales asset for the same reason. Audit trails, data residency and SOC 2 paperwork are procurement requirements now, and a solo operator who can produce them gets shortlisted against larger vendors who treat compliance as a delay. Security posture used to be the thing you fixed after the first enterprise deal. For a small AI product in 2026 it is often the reason the deal exists.
The honest verdict is that the category with the moat is rarely the category with the demo. The flashy consumer app gets attention and the boring workflow tool gets paid, and the founders who understand that difference stop chasing distribution on the strength of a model they do not control.
What the Median Product Actually Earns
Across a tracked set of more than 8,000 SaaS products, the average monthly recurring revenue is around $4,298 and the median is $145. Roughly 70 percent never clear $1,000 a month, and only about 6.1 percent reach $10,000. That gap is not a story about effort or talent. It's a distribution with a long tail, and the average is dragged upward by a handful of winners while the typical product sits quietly at the bottom of the curve.
The profitable middle looks different from both ends. The median profitable micro-SaaS runs about $4,200 in monthly recurring revenue at 64 to 76 percent margins, which is a modest, durable business rather than a rocket. Run the number backwards and it gets concrete: 35 accounts at $120 a month produces roughly $4,200 in revenue, which at 70 percent margin leaves close to $2,940 a month before your own time is paid for. That is reachable by one person with a narrow product and a sales list, and it is the outcome most first-time founders claim they want until they see how little it looks like a press release.
What kills that path is not competition from a bigger model. It's the assumption that margin will sort itself out once volume arrives, when volume is exactly what turns a mispriced plan into a compounding loss. The cost of a request does not fall because you grew. It falls because you routed it somewhere cheaper, cached what repeated, and priced the meter honestly.
December 1 is the useful test of all this. A rate card will change, a margin will move, and the founders who already know their cost per request will absorb it with a routing tweak while everyone else discovers it in a billing dashboard three weeks later. Profitability in AI software was never going to come from a better model layer that someone else owns and someone else prices. It comes from knowing what each customer costs you before they cost you the quarter, and almost nobody wants to do that work.
Strategic Takeaways
- ✓Inference is a variable cost that lands before payroll, so a per-seat price written in 2024 can turn a healthy product into a loss-making one without a single customer leaving.
- ✓The identical workload costs $120 or $1.20 per thousand requests depending on which model tier handles it, which makes routing a commercial decision rather than an engineering preference.
- ✓Vertical depth and a metered price protect a small software business far better than a clever prompt, because anyone with a coding assistant can clone a clever prompt by Sunday.

Maleek Joshua
Business & Work Editor at Largern, writing about markets, entrepreneurship, careers, and the scenes behind how companies and people make money.
Comments (0)
No comments yet
Be the first reader to share your perspective on this story.



