About Mozart AI
Mozart AI is the AI music generator built for artists. Its flagship experience, Vibe Sessions, puts an AI co-producer in the browser: describe the genre, mood, tempo, and style you want, and the agent composes, arranges, and refines the track through natural conversation. When artists want hands-on control, they switch to Studio, a full online DAW with piano roll, drum patterns, and mixing, with commercial rights included on everything they create.
Behind the conversational surface is a heavyweight agentic backend. Nearly all of Mozart's LLM traffic is non-streaming backend work, and 40% of requests demand structured JSON output: the agent plans arrangements, edits compositions, and calls tools programmatically.
The Challenge
The model landscape moves fast, and so does Mozart. The best model for the co-producer changes every few months, different tasks suit different models, and every new frontier release is a chance to make the product better. Building on provider APIs directly would have made that agility impossible:
- Every model switch is a migration. New provider SDKs, new auth, new error handling, regression testing, redeployment. Trying a new model should take minutes, not a sprint.
- Different tasks, different models. Composition, lightweight classification, vision analysis, and reasoning each have a different best model, often from a different provider.
- Single-provider fragility. Roughly 4-5% of first-attempt requests fail in any given month. Without fallback, every one of those is a broken session for an artist mid-song.
- Token economics at scale. Agentic music production is context-heavy. With billions of input tokens per month, paying full price for repeated context would crush margins.
- Scaling blind. Volume grew 12x in under a year. The team needed per-model, per-provider cost visibility to scale without surprises.
“The best model changes every couple of months. We wanted switching to be a config change, not an engineering project, without ever giving up reliability.”