AI Cost Control for SaaS

Every SaaS founder I work with eventually asks the same question: "How do I keep AI from eating my gross margin?"

Five patterns separate teams that have AI-cost discipline from teams that don't.

1. Tier routing by task difficulty

Not every task needs GPT-4o or Claude Sonnet. Most don't.

• Cheap tier (Haiku, GPT-4o-mini, Gemini Flash): classification, formatting, simple extraction. $0.001–0.005 per call.
• Mid tier (GPT-4o, Sonnet): general user-facing chat, RAG synthesis. $0.01–0.05 per call.
• Reasoning tier (o1, o3, Opus): hard problems that need actual thought. $0.10–1.00 per call.

Code your features to route by task complexity, not "use the smartest model for everything".

Result: 50–80% cost reduction on most features.

2. Prompt caching aggressively

Anthropic and OpenAI both offer prompt caching. If your prompt has any repeated prefix — system prompt, RAG context, few-shot examples — cache it.

Anthropic gives you 90% off cached tokens. For RAG-heavy use cases, this is a 3–5x cost reduction immediately.

Pattern: put the variable user input at the END of the prompt; everything before it can be cached.

3. Per-user rate limits with budgets

Every feature has a per-user rate limit AND a per-user monthly budget. Both are enforced server-side.

When a user hits their limit, the feature gracefully tells them. They don't break your bank.

Default budgets I use:

• Free tier: $0.50/user/month
• Paid tier: $5–25/user/month depending on plan
• Enterprise: negotiated, with overage billing

4. Eval-driven model swapping

Without an eval harness, you can't safely move to cheaper models. With one, you can.

Build the eval. Run quarterly: "Can we move this feature to a cheaper model and stay within quality bar?"

Often yes. The market keeps shipping cheaper-equivalent models.

5. RAG over fine-tuning, fine-tuning over big-context

For knowledge-heavy use cases, fight the urge to dump everything into context. RAG is cheaper at scale.

For format/style use cases, fine-tune a small model. Cheaper than calling Sonnet for every request.

The math founders miss

If your AI feature costs $0.05/use × 10 uses/user/day × 30 days = $15/user/month.

If your SaaS is $20/month, you have $5 left for everything else (servers, support, marketing, profit). That's tight.

If your AI feature costs $0.10/use, you're losing money.

Calculate this before launching the feature. Most founders don't.

Live patterns from my products

• Maeum Campfire: prompt caching for character context. 4x cost reduction.
• MAEUM Learn: tier routing across 10 LLM providers. Free tier sustainable at $0/user.
• GC OS: cheap tier for routine extraction, Sonnet for complex AI assistant queries.

What I build into every AI integration

Every AI feature I ship includes:

• Per-user rate limit + monthly budget
• Tier routing logic
• Prompt caching where applicable
• Cost dashboard (per feature, per user)
• Eval harness for quality regression detection

Email [email protected] to scope yours.