OpenAI shipped Presence on July 22, 2026, and it turned a fuzzy question into a clean fork. You can now get an OpenAI voice agent two ways: buy the managed platform, or build on the self-service Realtime API and AgentKit. The trap is treating the newer option as the better one. It isn't better or worse — it's a different debt, and which debt you want depends on what you're optimizing for.

Here's the decision in one screen, then the reasoning.

The fast answer#

Both run the same models underneath. This is not a quality decision. It's a decision about who owns the loop — the guardrails, the evals, the improvement cycle.

What each path costs#

The build path prices in the open, which is its quiet advantage. The Realtime API lists ~$0.05 per conversation minute on gpt-realtime-2.1 and ~$0.016 per minute on the mini tier (third-party math on the published rates). You can put that in a spreadsheet before you write a line of code.

The buy paths don't. Presence is enterprise contact-sales — no disclosed pricing, no public contract terms, delivered through OpenAI Forward Deployed Engineers. For reference on where managed CX agents land, bundled voice platforms typically run $0.07–$0.31 per minute, and a vertical like Decagon is estimated at roughly a $50k annual platform fee plus per-resolution fees, with contracts reportedly from ~$95k to $590k+ and a ~6-week onboarding. Sierra, Decagon, and Ada don't publish prices, so treat those as third-party estimates — but the direction is clear: managed platforms cost multiples of raw inference because you're buying the assembly, not the tokens.

When the vendor won't show you a price, you're not buying a product. You're buying a relationship — and pricing it is part of the diligence.

The three questions that actually decide it#

1. Is this agent the product, or a chore you want gone? If the voice agent is a support desk you'd rather not staff, buying is rational — the managed guardrails, approved actions, evals, and Codex-powered improvement loop are real work you'd otherwise build and hire for. If the agent is the thing customers pay you for, buying a deployment your competitor can also buy is buying a commodity. Own the loop. This is the same build-vs-rent logic we applied to auth: rent the undifferentiated, own the defensible.

2. Do you have the engineers? Presence exists precisely for teams that don't. The self-service stack assumes you can wire the Realtime API, connect your systems, and stand up your own evals — the shape of that work is what voice-agent stacks get you into. No voice-ops muscle, and the build path's open pricing is a mirage; the cost just moves into salaries and calendar time.

3. How much does portability matter? Presence locks you to OpenAI's models and roadmap, and the governance loop lives in their platform, not your repo. For a support line, fine. For a core capability, that lock-in is the expensive part — the day you want to route cheap calls to a smaller or open-weight model, the managed platform is exactly where you can't.

The honest recommendation#

For most solo founders and small teams, the Realtime API is the default — open pricing, model portability, and the agent stays yours. Reach for Presence when the agent is mission-critical CX, you lack the staff to run it safely, and a few weeks of white-glove setup is cheaper than the alternative. And whichever you pick, steal Presence's checklist — SOPs, guardrails, approved actions, evals, a feedback loop — because that list is now the definition of "production-ready," whether you buy it or build it.