Stage 1: prototype
- Use strong models freely; your volume is tiny and you're learning what works.
- Use one key with a small monthly limit.
- Write down your test examples as you go; they become your eval set later.
Stage 2: beta users
- Estimate cost per user action (a message, a summary, an image) from Usage Logs (estimating costs).
- Add per-user limits in your app: free users get a daily allowance.
- Separate keys for production and development.
Stage 3: launch and growth
- Make sure price per user ≥ AI cost per user × 3 or so, to cover other costs and refunds.
- Move high-volume, simple tasks to smaller models.
- Route hard requests to bigger models only when needed (routing).
- Cache repeated results.
Watch these numbers monthly
- AI cost per active user
- AI cost as a percentage of revenue
- The top 3 features by AI cost
Free tiers without bankruptcy
If you offer a free tier: use your cheapest acceptable model for it, set strict daily limits, require sign-up, and add basic abuse protection (rate limits, CAPTCHA on sign-up).
Get help
Building something genuinely useful? The Developer Partner Program sponsors promising projects with AI credits.
Claude, GPT, Gemini, DeepSeek, image and video models through one OpenAI- and Anthropic-compatible endpoint. Prepaid, pay only for what you use.