The real tell isn't that someone built a voice-driven murder mystery game — it's that running it overnight could bankrupt them.
The Summary
- A developer built a voice-driven murder mystery game using OpenAI's gpt-realtime-2.1 over WebRTC, where players interrogate AI suspects with their voice
- The economics are brutal: the model is so expensive the developer had to impose auth walls and 30-minute session limits to avoid going broke while sleeping
- Each suspect uses function calling to detect accusations, then a separate GPT-4o-mini judge evaluates whether the player actually presented required evidence (paraphrasing counts, fishing doesn't)
The Signal
This is a perfect snapshot of where voice agents are in 2026: technically impressive, economically terrifying. The developer originally built this 2-3 years ago when voice AI was "very early" but shelved it because the tech wasn't good enough. Now it works beautifully. The problem is the bill.
OpenAI's gpt-realtime-2.1 delivers speech-to-speech conversations over WebRTC with low enough latency to feel natural. But "natural" costs money at a rate that makes the creator anxious about overnight traffic. The solution: Clerk authentication (to track usage per user), a hard 30-minute timer, and presumably a lot of budget alerts.
"I really don't want to go broke while I sleep tonight."
The architecture reveals how builders are navigating AI economics in real time. The murder mystery uses two models in tandem: the expensive realtime model for the voice conversation, and GPT-4o-mini as a judge to evaluate evidence. This is agent scaffolding 101. Use the cheapest model that can do the job for each discrete task.
The judge mechanism is particularly smart:
- Suspects call a tool when they detect a direct accusation
- The tool captures who you accused and what evidence you stated
- GPT-4o-mini evaluates whether you genuinely presented the required evidence facts
- Paraphrasing counts, vague suspicion and fishing don't
This isn't just a game mechanic. It's a pattern for constraining agent behavior and measuring output quality without burning tokens on the expensive model. The realtime agent handles the conversational heavy lifting. The mini model handles structured evaluation.
The Implication
Voice agents work now. The compute costs are the bottleneck, not the capability. Builders are getting creative with hybrid architectures, usage caps, and auth walls, but these are stopgaps. The real question is how long before voice compute becomes cheap enough to run at scale without existential budget anxiety. Until then, every voice agent project will have a "please don't use this too much" disclaimer baked in. Watch for the first developer who figures out how to make voice interactions economically sustainable without sacrificing quality.