13 min read

How much does it cost to ship an AI-built app in 2026?

If you're a solo founder budgeting for your first AI-built SaaS, this is the honest cost breakdown — every line item, every free tier, every gotcha that turns a $50/month plan into a $1,200 surprise. Written for the moment before you pick your stack.

You're about to pick a stack. You want to know what it actually costs to ship and run a SaaS in 2026.

The marketing pages of every vendor will tell you "free tier covers most projects" or "starts at $20/mo." Reality is somewhere between $0 and $1,200/month for a small SaaS, and the variance is mostly about gotchas you only discover after the bill arrives.

This is the honest breakdown. Every line item. The free tier limits. The 2 AM bill-shock patterns.

TL;DR — what a solo-founder SaaS actually costs

| Stage | Monthly cost | What you're paying for | |-------|-------------|----------------------| | Pre-launch (building) | $0–50 | Mostly free tiers; $20/mo Cursor or similar AI tool | | Launch with first 100 users | $50–200 | Hosting + database + payments + email | | 100-1,000 paying users | $200–800 | Scaling tiers + monitoring + customer-support tooling | | 1K-10K users | $800–3,000+ | Higher-tier infra, multi-region, support team |

The variance at every stage is mostly: which AI APIs you call, how chatty your users are, whether you set spending limits, whether someone leaks a key.

Hosting + frontend (Vercel / Netlify / Cloudflare Pages)

Free tier (Vercel Hobby, Netlify Starter, Cloudflare Pages free):

  • Static hosting: unlimited
  • Serverless function invocations: 100,000-1,000,000/month depending on platform
  • Edge function invocations: similar caps
  • Bandwidth: 100GB-1TB/month free tier

Paid (Vercel Pro $20/seat/mo):

  • Higher invocation limits
  • Better caching, faster cold starts
  • Better support
  • Required for commercial use on most platforms

Gotcha — function execution time: Vercel Hobby caps function invocations to 10 seconds. If you call OpenAI for a 30-second response, your function times out and the user sees an error. Solution: streaming responses (streamed from your function back to the user as the LLM produces tokens) or upgrade to Pro for 60s + serverless function execution.

Gotcha — bandwidth: Static files are cheap; user uploads, generated images, and video can blow through bandwidth quotas fast. Watch the Vercel dashboard's "Bandwidth" panel weekly.

Realistic monthly cost:

  • Pre-launch: $0
  • 100 users: $0-20
  • 1K users: $20-60
  • 10K users: $60-200

Database (Supabase / Firebase / Neon / PlanetScale)

Supabase free:

  • 500MB Postgres database
  • 1GB file storage
  • 50K monthly active auth users
  • 50MB egress/month

Supabase Pro ($25/mo):

  • 8GB database (then $0.125/GB/mo)
  • 100GB file storage
  • 100K monthly auth users
  • 250GB egress

Firebase free (Spark):

  • 50K Firestore reads/day, 20K writes/day
  • 1GB storage
  • 50K auth users

Firebase Blaze (pay-as-you-go):

  • $0.06 per 100K reads, $0.18 per 100K writes
  • $0.18/GB stored, $0.12/GB downloaded

Gotcha — Firebase reads multiply with users. A typical SaaS dashboard does ~10 Firestore reads per page load. 1,000 daily active users × 10 reads × 30 days = 300K reads/day. That's 6x the free-tier daily cap. The bill goes from $0 to ~$50/month overnight when you cross the threshold.

Gotcha — egress bandwidth. Both Supabase and Firebase charge for data egress. Apps that serve large JSON payloads or many small files can hit egress caps fast. Monitor.

Realistic monthly cost:

  • Pre-launch: $0
  • 100 users: $0-25
  • 1K users: $25-75
  • 10K users: $75-300

AI / LLM API costs (the killer)

This is where vibe-coded SaaS budgets explode. Real numbers as of mid-2026:

OpenAI:

  • gpt-4-turbo: $10/M input tokens, $30/M output tokens
  • gpt-4o: $5/M input, $15/M output
  • gpt-3.5-turbo: $0.50/M input, $1.50/M output
  • gpt-5-nano (cheap tier): ~$0.15/M input, $0.60/M output

Anthropic Claude:

  • Sonnet 4.6: $3/M input, $15/M output
  • Haiku 4.5: $0.80/M input, $4/M output

OSS via DeepInfra / OpenRouter:

  • DeepSeek V3.2: $0.14-0.28/M input, $0.42/M output
  • GLM-5.1: similar
  • Llama 4 70B: ~$0.50/M input

Realistic per-user-month for a chatbot SaaS:

If a user sends 50 messages/month and each conversation has ~2K tokens of context + 500 tokens output:

  • OpenAI gpt-4o: 50 × (2K input + 500 output) = 100K input + 25K output = $0.50 + $0.375 = ~$0.88/user/month
  • Anthropic Sonnet 4.6: 50 × (2K + 500) = $0.30 + $0.375 = ~$0.68/user/month
  • DeepInfra DeepSeek V3.2: 50 × (2K + 500) = ~$0.04/user/month

For a chatbot SaaS at $20/month/user, OpenAI gpt-4o cost is ~4% of revenue. DeepSeek V3.2 cost is ~0.2%. The model choice changes your unit economics by 20x.

Gotcha — context window inflation. Long conversations, RAG retrieval, and chained tool calls compound. A user who sends 50 messages with full conversation history can use 1M tokens / month, not 100K. The "innocuous" chatbot SaaS hits $5-15/user/month in API costs at scale, sometimes exceeding the user's subscription price.

Gotcha — leaked API keys. A leaked OpenAI key gets discovered by an attacker's bot within 4 hours of any public commit. The bot uses your key to run cheap inference for someone else's product. By the time you wake up, the bill is $4,000.

The defense: spending limits at the provider (hard cap = 5x your expected monthly), automated secret scanning on every PR, and ideally a cost-firewall that throttles when a per-user spend anomaly fires.

Realistic monthly cost:

  • Pre-launch: $0-20
  • 100 users: $30-200
  • 1K users: $300-2,000
  • 10K users: $3,000-20,000

The 100x variance reflects model choice + caching + how chatty your app is.

Payments (Stripe / Paddle / Lemon Squeezy)

Stripe: 2.9% + $0.30 per successful charge. Free for failed charges. No monthly fees.

Paddle: 5% + $0.50 per charge. Higher rate but handles tax (sales tax / VAT) for you, which is real value for international SaaS.

Lemon Squeezy: 5% + $0.50 — similar to Paddle, also handles tax + acts as Merchant of Record.

Gotcha — chargebacks. Each chargeback costs ~$15 plus the disputed amount, even if you win the dispute. SaaS chargeback rate is typically <1%, but if it gets above 1% Stripe / Paddle starts threatening to drop you.

Gotcha — sales tax. US sales tax has 11,000+ taxing jurisdictions. EU VAT requires registration in 27 countries. Selling to international customers with Stripe alone (vs Paddle/Lemon) means you eventually owe back-taxes you didn't know you collected.

Realistic monthly cost: percentage of revenue. At $5K MRR with Stripe, ~$155/mo in fees.

Email (Resend / SendGrid / Postmark)

Resend: Free tier 3K/month, $20/mo for 50K, modern API.

SendGrid: Free 100/day, $20/mo for 50K. Older but everywhere.

Postmark: $15/mo for 10K transactional emails. Most reliable for transactional (verification, password reset, notifications).

Gotcha — sender reputation. A new domain sending email blasts gets flagged as spam. Warming up a domain takes 4-8 weeks. Solution: send transactional only at first, or use Resend/Postmark which manage reputation collectively.

Realistic monthly cost: $0-20 pre-launch, $20-50 with 1K users, $50-150 with 10K.

Auth (Supabase Auth / Clerk / NextAuth / Auth0)

Supabase Auth: included in Supabase pricing.

Clerk: free up to 10K MAU, $25/mo + $0.02/MAU after.

NextAuth: free, but you host yourself.

Auth0: legacy choice, expensive past free tier ($240/mo for 10K MAU).

Realistic monthly cost: $0-25 if Supabase Auth or NextAuth; $25-300 if Clerk or Auth0 at scale.

Monitoring + observability (Sentry / Better Stack / LogTail)

Sentry: free tier 5K errors/month, $26/mo for 50K.

Better Stack: free tier with full features for hobby; paid tier from $24/mo.

LogTail / Logtail: free 1GB/month, $24+/mo for paid.

Gotcha — error rate spikes. A bug that throws a frontend error in a useEffect can produce 1M errors in an hour. You hit your error quota in minutes. Set rate limits in the SDK (ignore-errors-after-N).

Realistic monthly cost: $0-26 most of the time; surprise spikes to $100+ during incidents.

Security tooling (Securie / etc.)

This is where I'm biased.

Securie: - Free: $0, 1 repo, 20 scans/month, 3 Day-1 specialists, show-only fixes - Indie: $12/mo, 3 repos, 100 scans/month, all specialists, auto-fix, cost-firewall, <15min CVE SLA - Solo Founder: $49/mo, 10 repos, 300 scans/mo, human concierge, production-readiness cert - Startup: $299/mo, 50 repos, 2000 scans/mo, 2 swarms/mo, SSO, audit-log export, SOC 2 evidence pack

The security-tooling alternative is "$0 because I don't have any" + the implicit cost of the first incident. Solo founders launching today without automated security review ship the canonical bugs at high frequency; the cost of one Hacker News post tweeting your data leak is unbounded.

Realistic monthly cost: $0-49 for solo founders; $299 once you have paying enterprise prospects.

Putting it together

Realistic budget for a 6-month-old, post-launch SaaS with 500 paying users at $20/mo (so $10K MRR):

  • Vercel Pro: $20
  • Supabase Pro: $25
  • LLM APIs (mid-tier): $400
  • Stripe fees: $310
  • Resend: $20
  • Clerk: ~$30
  • Sentry: $26
  • Securie Solo Founder: $49

Total: ~$880/month, ~9% of MRR.

Compare to: paying yourself $0 for the first 6 months.

Where the money actually goes

Two surprises, in order:

1. LLM API costs, if your app is AI-heavy. Pick your model carefully. DeepSeek V3.2 vs OpenAI gpt-4o is a 20x cost difference for similar quality on many tasks. 2. Stripe fees, once revenue is real. 3% feels small; on $10K MRR it's $300/month, which matters at solo-founder scale. Higher pricing tiers reduce the percentage friction.

Everything else is in the noise.

The cost mistakes to avoid

  • Not setting spending limits on every API. The default "no limit" is the worst possible default for solo founders.
  • Production secrets in preview deploys. Preview URLs are public; production Stripe + production database in preview = production charges from random PRs.
  • Not monitoring API spend. You should be checking the OpenAI / Anthropic / AWS dashboards weekly. Set alerts if they exist.
  • Picking gpt-4-turbo when gpt-4o or DeepSeek V3.2 would have worked. Most chatbot tasks don't need the highest-tier model. Test cheaper models early.
  • Forgetting about leaked keys. Every solo founder leaks at least one key in their first 18 months. Set up secret scanning + auto-rotation BEFORE the first leak, not after.

Related

Related posts