WhoDunnitAI Voice Murder Mystery — Pricing Impact

WhoDunnitAI launched on Hacker News on August 10, 2026, with an unusually concrete use of realtime voice AI: players question murder suspects aloud, compare their stories, and accuse a killer using the evidence they gathered.

Creator Chase Myers says the interviews run speech-to-speech on OpenAI’s gpt-realtime-2.1 over WebRTC. When a player makes a direct accusation, the suspect calls a tool that records the accused person and the evidence actually stated. A separate gpt-5-mini judge decides whether that evidence satisfies the case’s required facts.

The live tables above use today’s pricing dataset rather than frozen numbers in this article. There is no announced WhoDunnitAI subscription or per-game price.

What launched

The first case is Death at Blackwood Manor: Edmund Blackwood has been killed with digitalis, five people were trapped at the estate by a storm, and each can be interviewed as a suspect. The site describes the experience as a voice-driven investigation rather than a menu-based visual novel.

Myers says the idea began as a proof of concept two to three years ago and was rebuilt as realtime voice technology matured. The rest of the disclosed stack is Next.js, MongoDB, and Clerk authentication.

Layer	Model or service	Role in the game	Main cost pressure
Live interview	GPT-Realtime-2.1 over WebRTC	Listens and responds as each suspect	Continuous audio input, generated speech, and repeated case context
Final accusation	GPT-5 mini	Checks stated evidence against required facts	Text input, reasoning output, and retries
Session control	Clerk	Ties conversations to an authenticated user	Account friction and abuse prevention
Game state	MongoDB	Stores case and session data	Reads, writes, retention, and scaling

Why the 30-minute cap matters

The creator explicitly called GPT-Realtime-2.1 expensive and said the game uses a 30-minute timer to control overnight spend. That is the clearest pricing signal in the launch: long, open-ended voice conversations can turn a playful demo into an uncapped infrastructure bill.

At the time of our check, the landing page showed a 37-minute typical solve time, longer than the disclosed cap. That tension may frustrate careful players and suggests the game still needs either shorter investigations, stricter turn design, or a sustainable paid allowance.

The architecture also shows smart cost separation. The premium realtime model handles the experience users notice; a cheaper text model evaluates the final evidence. Builders should copy that routing principle instead of assigning every background task to the voice model.

What this means for voice-agent builders

WhoDunnitAI turns a common enterprise pattern—realtime conversation plus tools—into a game mechanic. Each suspect needs a stable identity, hidden facts, consistent testimony, interruption handling, and a bounded tool call. The judge then needs a structured rubric that accepts paraphrases without rewarding vague guesses.

That is relevant beyond games. Interview practice, language learning, sales simulation, training, and customer research can use the same split between an immersive voice role and a lower-cost evaluator.

Before building, compare the current OpenAI pricing page with Google AI pricing and the broader voice API pricing table. Use the AI token calculator for the text judge, then budget audio separately. Our GPT-Realtime retail-agent cost analysis covers a production deployment of the same model family.

Who benefits—and who loses

Players who enjoy interrogation, improvisation, and spoken role-play get a format static dialogue trees cannot easily reproduce. Indie developers also gain a compact reference architecture for tool-using voice characters.

The weak spots are equally clear. Mandatory signup drew immediate HN pushback, while early commenters reported that the game would not connect or move past dispatch confirmation. AI-generated voices may also repel players who prefer human performance. Those reports are anecdotal, but they matter during a time-sensitive launch.

Developers evaluating a modular voice layer can benchmark ElevenLabs voice agents against the same suspect scripts and latency targets.

Affiliate disclosure: AI Pricing Guru may earn a commission from the sponsored ElevenLabs link at no extra cost to you. It does not affect this analysis.

What to do now

Players should treat WhoDunnitAI as an early public demo: use a working microphone, allow browser audio access, and enter with a plan to question all five suspects inside the session limit.

Builders should log cost per completed case, not cost per minute alone. Track connection success, useful turns, repeated context, accusations, judge retries, completion rate, and how often users hit the timer. If the experience cannot finish reliably within its cost boundary, the model choice is not yet economically viable.

Bottom line

WhoDunnitAI is a sharp demonstration of what realtime voice models now enable: free-form interrogation, consistent characters, tool-triggered accusations, and evidence-aware judging. It is also an honest reminder that voice inference needs hard limits.

The product’s next test is commercial rather than technical: can it keep the mystery open long enough to be satisfying without making each investigation too expensive to offer?

Sources: WhoDunnitAI, the creator’s Show HN post, OpenAI’s official GPT-Realtime-2.1 model card, and API pricing documentation. Product and discussion checked August 10, 2026.