Solo Claude SaaS Cost Diary: $4 in Month 1, $87 in Month 6
A one-person Claude-powered SaaS spent $4.23 on the Anthropic API in month 1 (with two paying users) and $87 in month 6 (with 240 users) β not 240Γ more, because routing 78% of calls to Haiku and turning on prompt caching at month 2 absorbed most of the growth. This is the honest month-by-month breakdown: what each month's bill bought, what mistake cost the most, and the three settings every solo builder should turn on before user #50. Numbers are dated 2026 and rounded to the nearest dollar.
Month-by-month cost ledger
| Month | Active users | Calls/day | Total API ($) | Average $/user | Note |
|---|---|---|---|---|---|
| 1 | 2 | ~10 | $4 | $2.00 | All Sonnet, no caching. Demoware. |
| 2 | 15 | ~120 | $19 | $1.27 | Turned on prompt caching. Immediate drop. |
| 3 | 40 | ~400 | $31 | $0.78 | Added Haiku routing for short classifications. |
| 4 | 90 | ~1,100 | $48 | $0.53 | Added batch API for nightly digest emails. |
| 5 | 160 | ~2,200 | $71 | $0.44 | Hit a bug: streaming retries doubled costs for 3 days. |
| 6 | 240 | ~3,400 | $87 | $0.36 | After fixing retries, cost per user kept falling. |
Six-month growth: 120Γ users, 22Γ cost. The compounding lever is the per-user cost line β from $2 to $0.36, almost all of it from three settings rather than from cheaper models.
What each month bought (honest tally)
Month 1 β Demoware ($4)
Two friends signed up. Sonnet everywhere, no caching, ~10 calls/day. The bill was $4 and felt "wow this is free." Mistake here would not have shown up yet.
Month 2 β Caching ($19, jumped 5Γ)
First real users (15). Without caching, the projected bill was about $50. Turned on cache_control: {type: "ephemeral"} on the 3,200-token system prompt; cache read at 10% of input price kicked in after the first call per 5-minute window. Result: input cost dropped ~80%, total bill was $19 instead of ~$50. The 1.28-reuse break-even from the prompt caching guide clears any time the same prompt is reused twice within 5 minutes.
Month 3 β Haiku routing ($31)
Realized that 60% of API calls were short classifications (intent detection, language ID, simple yes/no). Routed those to Haiku 4.5 at $1/$5 instead of Sonnet at $3/$15 β same quality on those tasks, 3Γ cheaper. The model selection decision tree was the input for the routing rule.
Month 4 β Batch API for non-interactive work ($48)
Nightly digest emails (one per active user) processed in batch overnight. Batch API = 50% off both input and output. About 240 emails/night = 14,400/month, would have been ~$24 at sync rates; cost $12 in batch.
Month 5 β A real bug ($71, spiked)
Streaming responses were timing out on slow networks and the client retried. Each retry = full output cost again. For three days the bill ran 80% high before I caught it. Fixed with: explicit retry budget, request idempotency key, and an alarm at 2Γ rolling 7-day average. The error handling production guide covers the exact retry+idempotency pattern.
Month 6 β Maintenance ($87)
Cost per user actually continued to fall. New users coming in were converting at the steady-state $0.36/user/month. Headline number is the new monthly run rate, not a problem.
The three settings that did 90% of the work
If you do nothing else before user #50, do these:
- Prompt caching on any system prompt over 1,000 tokens β break-even at 1.28 reuses; almost every real workload clears this. Free 80%+ savings on the input column.
- Haiku routing for sub-500-token classifications β Sonnet's power isn't needed for "is this email spam: yes/no". 3Γ cost savings on whichever portion of traffic this is, no quality loss.
- Cost alarm at 2Γ rolling average β catches retry bugs, prompt drift, abuse. Cheaper to set up than to debug a 5Γ weekend spike.
See the cost calculator for your own numbers and the Cost Optimization Masterclass ($59) for the full audit checklist.
Three mistakes that cost the most (in order)
- No caching at month 1 β every demo eaten at full input price. Cost: ~$30 over 4 weeks. Wasted entirely.
- All-Sonnet routing until month 3 β Haiku-eligible traffic charged at 3Γ the right rate. Cost: ~$18 over 2 months.
- No retry budget until month 5 β the bug burned ~$22 in 3 days. Most-painful per-day single line item.
Total avoidable: ~$70 over 6 months. The cost of a single dinner in San Francisco β and exactly the price of the Cost Optimization Masterclass, which documents all of these and more.
What I'd tell month-1 me
- Caching is free money. Turn it on the first day you have a stable system prompt.
- Track cost per user from day 1. A line going down is the most reassuring metric in a solo SaaS.
- Build the cost alarm before you need it. The 3 days of debugging the retry bug were the only days I lost real sleep over the API bill.
Frequently asked questions
Is $87/mo a lot for 240 users?
For a Claude-heavy product (every request hits the model), no β that's about $0.36/user/month. For comparison, a typical small-SaaS infra bill (hosting + db + email + monitoring) is $40-150/mo, so AI is sometimes the smallest line. For lighter-touch products (Claude only on certain user actions), expect 0.3-1Γ this rate. Run your own scenarios in the cost calculator.
What if I don't have caching opportunities yet?
The instant your system prompt stabilises (post first product iteration), caching pays off. Most solo builders hit this around week 4. Even one-off tools with a stable instruction template benefit.
When does it make sense to route to Opus 4.7?
When the task is genuinely hard reasoning (multi-step plans, complex code synthesis, long-document analysis). For 90% of solo-product surfaces, Sonnet is enough. See the model decision tree.
What about Claude Pro subscription instead of API?
For Claude Code dev work, Pro/Max subscription often makes sense alongside the API for your product. The two billing surfaces are independent. See Claude Max vs API.
Get the full audit checklist
Claude API Cost Optimization Masterclass ($59) β every setting in this article + the audit checklist a paying customer used to take their bill from $2,100 to $187/month (see case studies). 30-day refund.