The Decision That Determines Your Unit Economics

If you're building a production AI application in 2025 and you're using GPT-4o for everything, you're either very well funded or you haven't looked at your bill yet. Model selection isn't just a technical decision — it's a unit economics decision that determines whether your product is financially viable at scale.

At ARM Creative Solutions, we built ZehnOra — a mental health platform with AI at its core — and we spent significant time working out exactly which model to use for which task. Here's the complete breakdown of what we learned.

The Price Gap Is Larger Than You Think

As of 2025, GPT-4o costs significantly more per million tokens than GPT-4o-mini. In a high-volume SaaS application with thousands of daily active users, this difference compounds aggressively. A feature that costs $200/month at GPT-4o-mini pricing might cost $1,400/month at GPT-4o — for the same user volume, the same feature, just a different model.

"Unlimited AI" at GPT-4o pricing is financially catastrophic at scale. We learned this early and designed hard usage caps and tier-based model routing into ZehnOra's architecture from day one. It is non-negotiable.

Our Production Model Routing Map

After extensive testing across ZehnOra's feature set, here is exactly how we route requests between models in production:

GPT-4o-mini — High Volume, Structured Tasks

  • Journal summarisation — Users journal daily. Summaries are structured, context-constrained, and don't require deep reasoning. Mini handles this perfectly at a fraction of the cost.
  • Mood tagging — Classification task. Short input, categorical output. Mini is ideal.
  • Session summaries — Post-session AI synthesis for practitioner dashboards. Structured format, bounded context. Mini performs excellently.
  • AI homework suggestions — Template-based output for therapeutic homework. Structured, repeatable. Mini more than sufficient.
  • AI Companion — Free tier — Budget-conscious, lighter conversational support. Mini keeps costs controlled.

GPT-4o — Safety-Critical and Premium Experiences

  • Crisis detection and intervention — Non-negotiable. When identifying potential self-harm risk from ambiguous language, you use the best available model. A missed signal is not a unit economics problem — it's a human one.
  • Therapy-style AI companion — Paid tiers — Nuanced, emotionally intelligent conversation requires GPT-4o's deeper reasoning and empathy. Users paying for premium mental health support deserve the best model.
  • Therapist AI assistant — Clinically adjacent co-pilot for practitioners requires high accuracy and sophisticated reasoning. GPT-4o only.
  • Progress reports and clinical insights — When generating reports that influence therapeutic decisions, model quality is paramount.

The Rule We Derived

After months of production experience, we arrived at a simple framework:

  • Use GPT-4o-mini for anything that is: high volume, structured output, bounded context, non-safety-critical, or where a "good enough" answer is genuinely good enough.
  • Use GPT-4o for anything that is: safety-critical, emotionally nuanced, requires complex reasoning, part of a premium paid experience, or where the cost of a mediocre output is higher than the cost of the better model.
The question isn't "is GPT-4o better?" — it almost always is. The question is "is GPT-4o meaningfully better for this specific task, at this specific cost ratio, given what's at stake?"

Hard Usage Caps: Non-Negotiable Architecture

Regardless of which model you choose, hard per-user, per-tier usage caps are essential for any production AI application. In ZehnOra:

// Simplified tier AI limits FREE tier: 10 AI companion messages/month (GPT-4o-mini) CALM tier: 50 AI companion messages/month (GPT-4o-mini) FLOW tier: 200 AI companion messages/month (GPT-4o) THRIVE tier: 500 AI companion messages/month (GPT-4o) // Crisis detection: unlimited across ALL tiers, always GPT-4o

Crisis detection is the one feature that is deliberately exempt from usage caps. A user in distress should never see "You have used your AI messages for this month." That would be a design failure with potentially catastrophic consequences.

The Bottom Line

If you're building a production AI SaaS in 2025: design your model routing strategy before you write your first API call, not after your first billing shock. The cost difference between thoughtful model routing and "use GPT-4o for everything" is not marginal — it can be the difference between a sustainable business and an AI product that burns money faster than it can earn it.

Build the routing logic into your architecture from day one. Make it configurable. And never, ever put AI safety features behind a usage cap.