What Your Feature Actually Costs at Each Scale
A typical AI feature: 4K input / 1K output per request, 5 requests per user per day, 50% cacheable. Monthly cost at three scales (verified 2026-08-28 rates):
| Model | Per request | 100 users | 1,000 users | 10,000 users |
|---|---|---|---|---|
| GPT-5.6 Luna | $0.0020 | $0.30 | $3 | $30 |
| DeepSeek V4 Flash (off-peak) | $0.0015 | $0.23 | $2 | $23 |
| Gemini 3.7 Flash | $0.0120 | $1.80 | $18 | $180 |
| Claude Sonnet 5 | $0.0180 | $2.70 | $27 | $270 |
| GPT-5.6 Terra | $0.0200 | $3.00 | $30 | $300 |
| GPT-5.6 Sol | $0.0500 | $7.50 | $75 | $750 |
At 1,000 users the entire AI stack costs $2-30/month on budget tiers — less than the coffee that built it. Even at 10,000 users, a budget tier runs $30. The flagship trap is the only thing that turns this table scary: Sol at 10K users is 25× the Luna bill for a quality delta your users won't perceive on routine features.
Protecting Your Margin
The real indie threat isn't the per-request rate — it's the usage multiplier: one power user running 50 requests/day with 100K-token contexts on an agent loop costs as much as 1,000 normal users. Cap it: max tokens per request, max requests per user per day, and a fallback route when the expensive path triggers.
The Indie Cost Rules
Default to budget, escalate explicitly
Luna or V4 Flash off-peak for the default path; Gemini 3.7 Flash for anything user-facing that needs polish; a flagship only behind an explicit 'deep' action.
Cache the system prompt from day one
90-97% off the repeated prefix on most providers — one line of code, and it compounds from the first request.
Set a monthly cap and alerts
Provider alerts at 50%/90% of a monthly budget. For a side project, $50/month is a generous cap; for a product, $500.
Run the 10-minute monthly ritual
Dashboard glance → Bill Doctor CSV (if BYOK) → adjust. The wrapped recap makes it shareable and honest.
Re-check quarterly
Prices move ~40%/year. The model that was right in March may be obsolete by August — the August 2026 changes prove it.
Frequently Asked Questions
How much does an AI feature cost for an indie product?
At 1,000 users doing 5 requests/day with a 4K-in/1K-out shape: $5-10/month on a budget tier (DeepSeek V4 Flash, GPT-5.6 Luna), $25-40 on a balanced tier (Gemini 3.7 Flash, Claude Sonnet 5), and $45-70 on a flagship. AI features are cheap at indie scale — if you avoid flagship defaults.
What is the cheapest model that doesn't embarrass me?
GPT-5.6 Luna ($0.20/$1.20, coding 71.4) and DeepSeek V4 Flash off-peak ($0.22/$0.66). For conversational features, Gemini 3.7 Flash ($1.50/$6) is the quality floor most founders should set — 76.1 coding and instant latency.
How do I price my product so AI costs don't eat the margin?
Compute cost per active user (requests/day × per-request cost), then set the price floor at 5-10x that number. At 5 req/day on Luna that's ~$0.01/user/month — a $10/mo plan has 99%+ margin; on Sol the same usage is ~$0.09/user/month and still 99%.
When should I worry about AI costs as an indie?
When any single line exceeds your hosting bill, or when a user can trigger 10x the median request cost (long contexts, agent loops, retries). Set provider alerts at 50%/90% of a monthly cap and review once a month — ten minutes keeps surprises away.
What's the one AI cost habit that matters most?
Never let a flagship be the default. Default to a budget or balanced model, escalate explicitly, and re-check quarterly — the model that was right in March may be obsolete by August. This one habit keeps indie AI bills at 5-10% of what default-stack teams pay.