A browser-based murder mystery game that lets players interrogate AI suspects by speaking aloud hemorrhaged roughly $200 in OpenAI costs as usage spiked after launch. Chase Myers, a software engineer at Indeed who built WhoDunnitAI as a solo project, watched his side hustle rack up charges at roughly $5 every ten minutes after the game hit the front page of Hacker News in mid-August.
"I launched this at midnight… a massive storm hit us… OpenAI account failed to fund lol," Myers wrote in one comment thread as usage spiked to 217 voice-enabled players. The experience offers a window into the economic pressures facing indie developers who build atop expensive real-time AI infrastructure—particularly when a viral moment arrives without warning.
Myers scrambled to implement emergency cost controls mid-launch: slashing free session times to 15 minutes, adding a donation button, and rolling out a bring-your-own-key option that lets users supply their own OpenAI credentials. The game, which offers a single case titled "Death at Blackwood Manor," connects players to five AI-powered suspects through OpenAI's gpt-realtime-2.1 model over WebRTC. Players verbally interrogate characters about a poisoned patriarch, gather clues, and make accusations without typing a word.
Of the 217 users who tried the voice feature, roughly 52 percent hit the free-time cap before cracking the case, according to live metrics on the site. Only eight players—3.7 percent—successfully solved the mystery. Typical solve time hovered around 21 minutes. That combination of low solve rates and high engagement meant prolonged sessions, which translated directly into mounting inference costs.
"We are at $100 usage since posting this," Myers noted in one Hacker News comment. Shortly after, he added: "I will probably cap it at $500 for you guys." The generosity masked a harder question about whether real-time voice AI can sustain indie economics at all.
The interrogation mechanic itself relies on speech-to-speech conversations over WebRTC. When a player directly accuses a suspect, the game logs the accusation and stated evidence through a tool call. A separate "gpt-5-mini judge" then reviews whether the player genuinely presented the required evidence, Myers explained in the launch thread. The system allows paraphrasing; players don't need to recite exact phrases.
To head off potential hallucination risks, Myers gave all suspect agents a "shared knowledge base" aimed at preventing different AI characters from contradicting one another about case facts. How well that holds up under pressure from hundreds of concurrent interrogations remains an open experiment.
Myers lists no company entity for WhoDunnitAI. His personal site shows he's worked as a Software Engineer II at Indeed since December 2023, following prior roles at Level Data. He holds a computer science degree from Western Michigan University. The game's site credits it simply as "Authored by Chase Myers."
The project uses Next.js, MongoDB, and Clerk for authentication alongside the OpenAI voice models. Myers requires users to create accounts to prevent runaway costs, a constraint that drew mild friction in early feedback. The Show HN post accumulated 202 points and 80 comments, with user reactions ranging from "This is really fun, great work!" to concerns about API key security and one dismissive "AI Slop" remark.
Several commenters requested features Myers implemented in real time during launch weekend, including the bring-your-own-key option and a demo video—now running for 3 minutes and 32 seconds on the landing page. Users who supply their own OpenAI API key can play without time limits, with the key stored in the browser's localStorage and sent via an x-openai-key header to mint ephemeral client secrets for WebRTC.
WhoDunnitAI enters a small but active niche. Murder Mystery Mayhem AI launched on Steam Early Access in September 2024 with text and experimental voice conversations subject to weekly limits. Everyone's a Detective, a browser game where suspects answer open questions, ran a public alpha earlier this year. Open Verdict, which promises iOS and Android versions with open-question interrogation, remains in pre-launch. The Day She Died, a browser whodunnit with no fixed culprit, launched roughly three weeks before Myers's project.
Older comparables include St. Noire, a 2019 Alexa-hosted voice board game. A separate recent game jam project also used "Death at Blackwood Manor" as a case title, employing Deepgram's voice agent API with Claude—a similar mechanic but unrelated to Myers's work.
Myers has not disclosed pricing tiers, a content roadmap beyond Case File No. 001, or plans to incorporate. The donation button and bring-your-own-key model shift costs to users but sidestep the longer question of sustainable monetization. His comment during the launch surge—"I will probably cap it at $500 for you guys"—signals both generosity and uncertainty about the economics of real-time voice AI for solo builders operating without venture backing or reserves to absorb sudden traffic.
Whether WhoDunnitAI evolves into a business or remains a curiosity depends partly on whether Myers can find a pricing structure that works. For now, the game stands as both a technical achievement and a cautionary tale: build something compelling with cutting-edge AI, and the costs can spiral faster than a murder plot.
