The AI Micro-SaaS Unit Economics Report 2026

Every "AI is a feature, not a product" take and every "wrappers have no moat" thread argues the same point from the same missing data: nobody publishes the actual unit economics. This report computes them. It runs one reference AI product through the hub's cost, margin, vector-DB, support, pricing, and payback engines, holds every input constant, and varies only the model tier and the user scale. Every number below is rendered live from a shipped engine bundle at build time and independently recomputed in continuous integration, so the prose and the math cannot diverge. Model prices are the list rates in the hub's model catalog, verified against the vendors' pricing pages on 2026-10-09; the hosting, auth and monitoring prices were re-checked on their pricing pages on 2026-10-19.

1. The question this report answers

The 2026 question is not "is AI useful" — it is "can a solo founder or small team run an AI product at a margin that pays a salary and funds acquisition." That decomposes into four sub-questions this report answers with engine output: what is the gross margin per user by model tier, what is the full-stack monthly bill at 100 / 1,000 / 10,000 / 100,000 users, what do the supporting lines (vector DB, AI support) actually cost, and what price and CAC payback the economics require. The answer is tier-dependent and, for the cheaper tiers, more favorable than the discourse assumes.

2. The reference product and its inputs

The reference product is a single-seat AI SaaS at $39/month. Each user makes 10 model API calls per day at 2,000 input tokens and 600 output tokens per call — a realistic profile for a retrieval-augmented assistant that stuffs context into the prompt. The infrastructure stack is Vercel Pro ($20), Supabase Pro ($25), Clerk auth (free under 50,000 monthly retained users), Resend email (free tier), Sentry Team monitoring ($26), a $12/year domain, and $50/month of "other" that absorbs the vector DB and miscellaneous SaaS. These inputs are held constant across the entire report; only the model tier and the user count change. Holding everything else fixed is what isolates the variable that actually moves the economics.

3. Margin by model tier at $39/month

The AI Product Margin engine takes the per-user usage profile and the model's input/output price and returns the gross margin per user and at scale. The only input that changes between the three runs below is the pair of token prices.

3a. Premium tier (GPT-5.5, $5/$30)

GPT-5.5 lists at $5 per million input tokens and $30 per million output tokens^[1]. At 10 calls/day, 2,000/600 tokens, plus $0.20/user hosting and $0.30/user other costs, the engine returns an AI cost of $8.40/user/month, a total per-user cost of $8.90, and a 77.2% gross margin. AI API is 94.4% of the per-user cost. At 10,000 users that is $84,000/month on the model alone against $390,000 of revenue — still a 77.2% margin business, but one with real cost of goods sold that a thin pass-through cannot wish away.

3b. Mid tier (Claude Sonnet 5.5, $2/$10)

Claude Sonnet 5.5 lists at $2 input and $10 output per million tokens^[2]. Holding every other input identical, the engine returns an AI cost of $3.00/user/month, a total per-user cost of $3.50, and a 91.0% gross margin. The model bill drops from $8.40 to $3.00, a 2.8x reduction, for the same call volume. AI API is still 85.7% of the per-user cost. At 10,000 users the model line is $30,000/month, leaving $355,000 of gross profit on $390,000 of revenue.

3c. Budget tier (Claude Haiku 5.5, $0.10/$0.50)

Claude Haiku 5.5 lists at $0.10 input and $0.50 output per million tokens for prompts up to 100,000 tokens^[2]. The engine returns an AI cost of $0.15/user/month, a total per-user cost of $0.65, and a 98.3% gross margin. At this tier the model API is only 23.1% of the per-user cost — the $0.30/user "other" line is now the dominant cost driver. A budget-tier AI product is structurally a software business: its COGS is so low that it behaves like a pure-SaaS margin profile and survives a model price war without flinching.

The margin ladder is the headline finding: 77.2% premium, 91.0% mid, 98.3% budget. The 2.8x difference in model cost between premium and mid is worth about 14 margin points; the further drop to budget adds another 7 points and flips the dominant cost driver away from the model entirely. The strategic read: choose the cheapest model that clears your quality bar, because every tier down is margin you keep.

4. The full-stack monthly bill at four user tiers

Margin per user answers the per-unit question; the AI Stack Cost engine answers the total-bill question by projecting the entire stack across 100, 1,000, 10,000, and 100,000 users. The run below is the premium tier (GPT-5.5 priced via the custom-model input at $5/$30) on the reference stack.

The premium stack returns $962/month at 100 users, $8,522 at 1,000, $84,122 at 10,000, and $841,147 at 100,000. AI API is 99.85% of the bill at 10,000 users. The only non-AI line that ever turns on is Clerk auth, which crosses its 50,000-user free ceiling and bills $1,025/month at 100,000 users ($25 Pro base plus $0.02 for each of the 50,000 extra users)^[5]. Everything else — $20 Vercel Pro^[6], $25 Supabase Pro^[4], $26 Sentry, $1 domain, $50 other — is flat from the first tier to the last. The mid tier runs the same stack at $30,122/month and the budget tier at $1,622/month at 10,000 users; the spread is set entirely by the model line. Price your own stack and model in the AI Stack Cost Calculator.

5. The vector DB line: RAG vs self-host

A retrieval-augmented product needs a vector store, and the discourse treats it as a major cost center. It need not be. The Embeddings DB Cost engine prices 2 million 1,536-dimension vectors with 30,000 queries and 3,000 ingests a day across four options.

At this scale the engine returns LanceDB on Cloudflare R2 object storage at $0.94/month and Turbopuffer at $16.67/month, just above its $16 Launch minimum, for a 14.31 GB stored footprint. Pinecone Standard comes to $212.87/month^[3]: Pinecone bills a query 1 read unit per GB of namespace, so each of the 900,000 monthly queries costs about 14 read units at $16 per million^[10], and reads alone run past $200 (the Pinecone vs Qdrant pricing breakdown shows where a free Qdrant cluster undercuts it entirely). Postgres with pgvector on Supabase comes to $977.58/month, because Supabase's own pgvector sizing puts an in-memory HNSW index of 2 million such vectors on its 64 GB 4XL compute at $960 a month^[11][4]; on a Postgres you already run with memory to spare, the same vectors cost close to nothing extra. Which option you pick decides whether the vector DB hides inside the reference stack's $50 "other" allowance or surfaces as its own line: Turbopuffer ($16.67) and LanceDB ($0.94) fit inside the allowance, Pinecone's serverless reads make it a separate line of about $213, and a dedicated pgvector database one of about $978. Pinecone's line is about 13% of the budget-tier stack's $1,622 a month and about 0.7% of the mid tier's $30,122. The RAG vs fine-tune total cost spoke works through the architecture choice in full.

6. The AI support line: deflection economics

The last operating line worth modeling is support. The AI vs Human Support Cost engine compares an AI-first desk against human-only at 3,000 tickets a month, 10 minutes per human ticket, a $30/hour loaded cost, 5,000 tokens per AI-resolved ticket on Claude Haiku 5.5 ($0.10/$0.50), and a 4-minute escalation overhead. The single biggest lever is the deflection rate.

At 70% deflection the engine returns a human-only cost of $15,000/month against an AI-first cost of $6,304.50/month, a $8,695.50 monthly saving, 58.0% lower, at $2.10 per ticket versus $5.00. The token cost of resolution is trivial; the saving is entirely in deflected human minutes. At 40% deflection the saving collapses to $2,395.50/month (16.0%), because the escalated 60% still carries full human handling plus the escalation overhead. Deflection rate, not token price, is the variable that decides whether AI support pays. The support break-even spoke charts the full deflection-versus-volume surface.

7. Break-even price and CAC payback

The economics only matter if the price clears them. The Micro-SaaS Pricing engine takes the premium-tier per-user cost ($8.40 API plus a small fixed allocation), a target 80% gross margin, and a $19–$49 competitor band, and returns the price the margin requires.

The engine returns a price floor of $42.50 and a suggested price of $42.50 to hold 80% gross margin, with a value-justified ceiling of $58.80 at an 85.6% margin. The reference $39 price sits just below the 80%-margin floor — which is why the premium tier lands at 77.2% rather than 80%. The fix is either a $3.50 price increase or a one-tier-cheaper model. Pairing the $39 price with a $180 CAC, the CAC Payback engine below returns a 6-month payback, a 24-month LTV of $722.64, and a 4.01x LTV:CAC ratio, which the CAC payback calculator rates Excellent and Good on its own bands; for reference, David Skok's rules of thumb are to recover CAC within 12 months and keep LTV above 3x CAC.

8. The wrapper-viability verdict per architecture

The architecture choices in this report reduce to a small verdict matrix, each row backed by an engine number above:

• Budget-tier wrapper (Claude Haiku 5.5, $0.10/$0.50): viable and durable. At a 98.3% gross margin the model is no longer the dominant cost; the product behaves like pure SaaS and survives a price war. This is the safest architecture for a solo founder.
• Mid-tier wrapper (Claude Sonnet 5.5, $2/$10): viable with healthy headroom. A 91.0% margin leaves ample room for CAC, salary, and reinvestment. Most production AI products belong here: good enough quality, cheap enough COGS.
• Premium-tier wrapper (GPT-5.5, $5/$30): viable but defends a real COGS. A 77.2% margin is a genuine business, but it sits below the 80% pricing floor at $39, so a premium-model product must either charge $42.50+ or add workflow value a buyer cannot replicate by calling the API directly.
• RAG over object storage or an existing Postgres: free of the vendor floor. At $0.94 to $16.67/month (LanceDB, Turbopuffer), or near zero as pgvector on a Postgres you already run, the retrieval layer is never the binding constraint; it is the cheapest way to add defensibility (your data) without adding meaningful cost. A Postgres provisioned just to hold the index in memory is the exception, at about $978/month for 2 million vectors on Supabase.
• Thin premium pass-through with no retrieval and no workflow: the one fragile architecture. A premium model, no proprietary data, and no workflow lock-in is the only profile this report finds genuinely exposed — every input is a commodity the buyer can source themselves.

The overall verdict: you can make money running an AI product in 2026, and the binding decision is the model tier, not the hosting stack, the vector DB, or the support desk. Choose the cheapest model that clears your quality bar, price at or above the margin floor, and add data or workflow value the raw API cannot. Re-run every engine in this report with your own numbers — the inputs are visible in each block above — and verify the per-vendor list prices on the linked pricing pages before any decision^[7][8][9].

Frequently asked questions

Can you actually make money running an AI product in 2026?

Yes, but the model tier decides it. On a $39/month product making 10 API calls per user per day at 2,000 input and 600 output tokens, the AI Product Margin engine returns a 77.2% gross margin on a premium model (GPT-5.5 at $5/$30 per million tokens), 91.0% on a mid-tier model (Claude Sonnet 5.5 at $2/$10), and 98.3% on a budget model (Claude Haiku 5.5 at $0.10/$0.50). The product is viable at every tier; the cheaper the model, the more headroom you keep for CAC and the founder's salary.

What is the single biggest cost in an AI SaaS?

The model API, by a wide margin. On the premium-tier full stack the AI Stack Cost engine attributes 99.85% of the bill at 10,000 users to the model API line; even on the budget tier it is 92.5%. Hosting, database, auth, email, and monitoring stay on free or low-fixed tiers well past 10,000 users, so they are rounding error next to inference. Cost optimization in an AI product means prompt design, output caps, caching, and model routing, not hosting choice.

How much does an AI SaaS cost to run at 10,000 users?

At 10,000 users with 10 calls per user per day, the AI Stack Cost engine returns $84,122/month on the premium tier (GPT-5.5), $30,122/month on the mid tier (Claude Sonnet 5.5), and $1,622/month on the budget tier (Claude Haiku 5.5). The non-AI lines are identical across all three: $20 Vercel Pro, $25 Supabase Pro, $0 Clerk (well under its 50,000-user free ceiling), $26 Sentry Team, $1 domain, $50 other. The model choice moves the bill by more than 50x.

Is a vector database expensive for a solo AI founder?

Usually not, if you pick by billing model. At 2 million 1,536-dimension vectors with 30,000 queries a day, the Embeddings DB Cost engine returns LanceDB on object storage at $0.94/month and Turbopuffer at $16.67/month, just above its $16 minimum. Pinecone Standard comes to $212.87/month, because each query is billed about 14 read units on the 14.31 GB index, and a dedicated Postgres sized to hold pgvector's index in memory to $977.58 on Supabase. The cheaper options are a rounding error next to the model API line in every tier.

What break-even price does an AI SaaS need?

Lower than most founders fear. To hit a target 80% gross margin on the premium-tier per-user cost of $8.50, the Micro-SaaS Pricing engine returns a price floor of $42.50 and a value-justified ceiling of $58.80. At $39/month the premium tier still clears 77.2%, and the CAC Payback engine shows a $180 CAC recovering in 6 months with a 24-month LTV:CAC of 4.01x, which the CAC payback calculator rates Excellent on its own bands; David Skok's rule of thumb is to recover CAC within 12 months.

Are AI wrappers viable as a business in 2026?

Viability depends on the model tier and whether the wrapper adds workflow value beyond the raw API. A budget-tier wrapper at 98.3% margin is structurally a software business and survives a model price war; a premium-tier wrapper at 77.2% margin is viable but carries real COGS and must defend its price against a buyer who can call the same API. The thin pass-through wrapper with no retrieval, no workflow, and a premium model is the only architecture this report finds genuinely fragile.