Introduction

Taan, a Y Combinator startup, launched Taan v1-preview on November 12. The team calls it "the most controllable foundation music model." It's aimed at one problem most music generators still have: you ask for "violin at exactly the 11th second," the model ignores you, and you reroll until a take happens to fit.

Taan works the other way round. You can give it a video and it scores to the cuts and mood. You can also add a script or a reference track, or write plain-language timing like "8 seconds of tension, a 2-second pause, then 15 seconds of uplift." When most of a track is right, you ask it to change one instrument or one timing and leave the rest alone. It's a closed, hosted model. You can use it today in Taan's web studio, from an AI assistant over MCP, or through a REST API. The preview only covers orchestral and cinematic music so far.

Sources:

(November 12, 2026), taan.ai, the MCP guide, the Music API v1 reference, and Taan's TimingBench results and paper.

What Shipped

  • Name and maker: Taan v1-preview, from Taan (taan.ai), a company in YC's F26 batch.
  • Modality: instrumental music generation and editing. The output is a full mix in WAV (the default) or MP3, plus separate instrument stems when they're returned.
  • Inputs: a text prompt, or any mix of a video, a script and a reference audio track, plus written musical direction. The launch post also mentions humming a melody, naming exact notes, or handing it an image or a MIDI file. The current API docs don't list image or MIDI inputs, so I'd treat those as studio-side or coming later.
  • Timing control: you can say when instruments come in or drop out, where the silences go, and what the tempo is. With a video, Taan reads the cut timing and mood and syncs pauses and changes to the edit.
  • Partial edits: you can refine a saved version, for example "make the violin happier and move its entrance from 8 to 5 seconds," while the piano and drums stay where they were.
  • Open vs closed: closed weights, hosted only. Taan says it trained on 100% licensed music and gives you perpetual commercial rights to the output.
  • Preview limit: genres are limited to orchestral and cinematic, with more promised.

Where to Run It Today

  • Web studio: app.taan.ai. Upload a video (up to 500 MB on paid plans), add direction, and download the mix and stems.
  • MCP: add the server https://mcp.taan.ai/mcp to Claude Code, Codex, Cursor or VS Code, sign in in the browser, and ask for music in the chat. You can then refine the same version in later messages. Checking your balance and job progress is free.
  • REST API: the base URL is https://app.taan.ai, authenticated with a bearer key from the API keys page. The flow is POST /api/v1/music/intents, then optional signed uploads, then POST /api/v1/music/generate, then poll GET /api/v1/jobs/{job_id}. Edits go through POST /api/v1/music/refine with the music_version_id. Jobs typically take a few minutes. There's an OpenAPI file and a Python client on the docs page.
  • Pricing:
    • Creator: $15 a month for 25 minutes of music (15,000 credits).
    • Agency: $45 a month for 60 minutes (36,000 credits).
    • Enterprise: custom.
    • MCP/API access, commercial rights and stems are included on both paid plans. I didn't see a free tier on the pricing page.

Limits Worth Knowing

  • Duration: 10–240 seconds without video (the default is 30). With a video, it matches the video's length.
  • File inputs: one video per request (MP4, MOV or WebM, 250 MiB over the API), one reference track (MP3, WAV, FLAC, M4A, OGG or AAC, 50 MiB), and one script (TXT, PDF or DOCX, 20 MiB).
  • Text: prompts and musical direction are capped at 8,000 characters, and script text at 100,000.
  • Refinement: a refine job can come back changed, unchanged or unsupported. An unsupported edit returns your original version with an explanation rather than a new take.
  • API projects: these are separate from studio projects, and there's no list endpoint, so keep your project, job and version IDs.

Prompts That Steer It

Timing-first briefs in the style Taan documents:

  • Create a tense cinematic track for exactly 30 seconds. Bring in drums at exactly the 6th second.
  • Silence for the first 3 seconds, low strings enter at 3s, full orchestra hits at 12s, pause at 20s, gentle piano resolves to the end.
  • Orchestral score at exactly 92 BPM, warm and hopeful, no percussion.
  • Refinement: Keep the duration and every other instrument. Make the violin brighter and move its entrance from 8 seconds to 5 seconds.
  • Video: upload the cut, then add Swell into the reveal at the final cut; stay sparse under dialogue.

Numbers (Self-Reported)

These come from Taan's own TimingBench. It's 250 prompts asking for 30-second instrumental clips with 399 timed events and 50 tempo targets. Taan published the snapshot in August 2026 and doesn't say which checkpoint it tested.

  • Overall instruction success:
    • Taan: 96.1%
    • ElevenLabs v2: 28.9%
    • ElevenLabs v1: 25.4%
    • Google Lyria: 23.9%
    • Suno: 22.0%
  • Timed events within ±500 ms:
    • Taan: 91.3%
    • ElevenLabs v2: 12.8%
    • Google Lyria: 8.6%
    • ElevenLabs v1: 8.0%
    • Suno: 6.8%
  • Tempo accuracy:
    • ElevenLabs v1: 93.0%
    • Taan: 92.0%
    • Google Lyria: 89.7%
    • ElevenLabs v2: 87.0%
    • Suno: 86.7%
  • Audio fidelity (FAD, lower is better):
    • Google Lyria: 2.6
    • Suno: 3.1
    • ElevenLabs v1 and v2: 3.5 each
    • Taan: 4.8, the worst score in the table.

So on Taan's own numbers, it's far better at landing a cue on time and somewhat behind on raw audio fidelity.

Caveats

  • Everything above is vendor-run. Taan wrote the benchmark and scored itself on it. Its paper is candid about measurement limits, for example timing error floors of about 8 ms for drums up to 218 ms for bass.
  • It's orchestral and cinematic only for now. If you need pop vocals or electronic drops, look at ACE Music v1 or Suno's MIDI-start songs instead.
  • It's instrumental only, with no vocals mentioned anywhere.
  • There's no open-weight release, and the pricing page shows no free plan.

Conclusion

Taan v1-preview is a hosted, closed music model built for scoring picture rather than writing hit songs. It follows timed cues, scores to your cuts, and lets you fix one instrument without rerolling the rest. If you cut trailers, shorts or game cinematics and keep fighting generators over timing, try it from your coding assistant over MCP or the $15 Creator plan. Then check the fidelity yourself against Lyria or Suno on a real cue.

—Anabel ♡