You pick up your phone. The suspect is on screen — a nervous-looking AI chatbot. You ask, “Where were you at 9 PM?” Your voice, your tone, your timing. The AI stumbles, tries to bluff, but you catch the inconsistency. You feel a rush — you’re not just playing a game; you’re interrogating a machine that might break under pressure.
That’s the hook of WhodunnitAI, a voice-driven murder mystery built on OpenAI’s gpt-realtime-2.1. But don’t call it a game. This is a publicly hosted stress test for real-time voice AI — and the murder mystery is just the scaffolding.
The creator, a solo developer, has been wrestling with this project for years. The original version from 2023 was a proof of concept. Now, with voice AI finally mature enough, he’s launched it — but with a brutal constraint: every minute of player freedom is a minute of cloud inference cost. The result? A 30-minute timer, forced authentication, and a design that teeters on the edge of bankruptcy.
Here’s the architecture that turns this into a benchmark: Each suspect has a tool that captures your direct accusation and a faithful list of the evidence you actually stated. A separate GPT-5-mini judge then decides which required facts you genuinely presented. Paraphrasing counts. Vague suspicion and fishing don’t. The suspects are the test suite. The mystery is a benchmark in disguise.
Most people will evaluate this as a game — is it fun? Is the story good? But the real question is: Can the AI sustain a coherent, evidence-based interactive narrative under the pressure of a real human voice? That’s the tension. The creator is torn between enabling an open-ended, immersive experience and the brutal unit economics of gpt-realtime. He’s essentially running a live experiment on the limits of conversational AI, and you’re the subject.
I tried it. The first suspect I grilled — a butler with a shifty alibi — started cracking when I repeated my question with a different inflection. The AI’s response lagged, then recovered with a contradictory detail. I felt like a detective, but also like a QA tester. Every accusation I made was a test case for the model’s consistency.
If you’re building on frontier AI, this is a case study you can’t ignore. It shows what it takes to productize real-time voice models: cost control, user constraints, and a thin layer of logic (evidence extraction and judging) around a raw conversational model. This project is a wake-up call: the future of voice AI won’t be about open-ended chatbots — it’ll be about controlled, high-stakes interactions that demand precision.
So go ahead, make an account (I know, I know, but it’s worth it), and interrogate a suspect. But remember: the AI is the one being tested. And you’re the one who decides if it passes.
FAQ
Q: Is this actually a game, or is it just a tech demo?
A: It's both. The murder mystery is a fun, engaging wrapper, but the real purpose is to test the limits of real-time voice AI. The creator even admits the timer and authentication are cost-control measures, not game design choices.
Q: What's the practical implication for AI developers?
A: This project shows that building a real-time voice AI product is expensive and requires tight constraints. You need a separate evaluation layer (like the judge model) to ensure consistency, and you must design for cost from day one. The 30-minute timer is a feature, not a bug.
Q: Isn't the contrarian take that this is just a game with a gimmick?
A: The hot take: Most people will dismiss this as a novelty, but it's actually a blueprint for the next generation of AI applications. By hiding the benchmark inside a game, the creator gets real user behavior data that no lab test can replicate. The murder mystery is the Trojan horse for AI research.