OpenAI's Decisions API and the Model Routing Era: What App Builders Need to Know

OpenAI's new Decisions API constrains LLM outputs to a bounded list of answers — and combined with mature model routing infrastructure, it marks a shift toward practical, cost-predictable AI features in production mobile apps.

OpenAI unveiled the Decisions API at DevDay 2026 in late September, and while it launched in limited preview — with a broader release reportedly imminent — it signals something concrete for developers shipping AI-powered features in mobile apps. Paired with increasingly mature model routing infrastructure, it points toward a moment when adding AI to a production iOS or Android app stopped being a cost gamble and started being a predictable line item.

What the Decisions API Actually Does

The premise is deliberately narrow: you supply a question, a bounded list of allowed answers, and a piece of context (text or image). The API returns exactly one of your allowed answers. No free-form generation, no hallucinated options outside the set you defined — just structured classification backed by a frontier model.

For app builders, that constraint is the point. Use cases that become practical:

• Support routing: Classify an incoming ticket into one of five queues without a full chat response.
• Content moderation: Decide if a user review, photo, or comment is appropriate — without generating explanatory text you don't need.
• Feature routing: Given a user's free-text input, determine which AI workflow to invoke.
• Agent next-step selection: From a defined policy list, pick the action an in-app AI agent takes next.

The key caveat: as of early October, pricing has not been published and the API remains in limited preview. According to OpenAI's public statements at DevDay, broad availability is planned for the near term. If you're scoping an AI-heavy feature for Q4 2026, it's worth watching but premature to build a hard dependency on.

The Model Routing Era Is Here

The Decisions API lands in a market that has been quietly restructuring. By mid-2026, there are now three viable frontier tiers competing for every AI call — cheap-and-fast (Claude Haiku at $0.25/million input tokens, Gemini 1.5 Flash at $0.10/million, several open-source options below $0.05/million), balanced midrange (Claude Sonnet, GPT-4o mini), and heavyweight (Claude Opus 5, GPT-6 Luna, Gemini Pro, in the $5–15/million range). The cost spread between the cheapest and most powerful is 100x or more.

Intelligent routing tools — LiteLLM, OpenRouter, Portkey, and several newer managed services — sit between your app and AI providers, sending simple requests to cheaper models and complex ones to capable ones. Research cited by multiple teams suggests 30–70% cost reductions while maintaining output quality on most production workloads.

For a subscription productivity app adding AI writing assistance, this spread can mean the difference between $0.10 and $0.003 per user session — a gap that determines whether the feature is economically viable at all, especially across lower-ARPU territories where PPP-adjusted pricing already compresses margins.

Practical Implications for Q4 App Launches

If you're deciding whether to add AI features before the holiday window, these two trends — constrained-output APIs and routing maturity — suggest a few concrete approaches:

1. Start with bounded tasks. Moderation, routing, and classification are much easier to test, debug, and cost-model than open-ended generation. The Decisions API pattern formalizes what good developers were already doing manually: constrain the output space before touching a frontier model.
2. Don't use one model for everything. Even a basic two-tier setup (fast model for simple calls, powerful model for complex generation) cuts costs materially. Tools like LiteLLM make this a configuration change, not an architecture rewrite.
3. Revisit the on-device calculus. Apple Intelligence and on-device Core ML models are zero marginal cost, but capability-limited. Cloud models are more powerful but metered. A routing layer helps you decide where each call lands — and the gap between on-device and cloud is narrowing as on-device models improve. See the on-device vs. cloud AI cost breakdown for the economics.
4. Think regionally. AI API costs are billed in USD, but your effective per-user cost varies by territory once you factor in subscription ARPU and purchasing power parity pricing. A feature that's margin-positive in the US may be margin-negative in markets where your price point is 60–80% lower.

It's not yet clear whether Apple will integrate constrained-output AI classification natively into future App Store review tooling or Siri intent matching — but the direction is consistent with how both Apple and Google are building AI into their developer platforms.

Sources and Further Reading

• OpenAI — DevDay 2026 announcements and API documentation
• Google Developers Blog — Building the agentic future: I/O 2026 highlights
• Anthropic — Claude model tiers and pricing
• RevenueCat Blog — App economy benchmarks and subscription data