Fable 5 Wrote a Fugue

I asked Anthropic's new model for a four-voice organ fugue in G minor, at least five minutes long, following period conventions, with some way to play it back. What came out is below. Press play.

Player not loading? Open it in its own tab.

A few facts about what you're hearing. It runs about 5:10, 84 bars, four voices. The structure is the real thing: an exposition with a tonal answer and countersubject, episodes built from the subject's head motive, entries in B-flat major and C and D minor, the subject in inversion, two strettos, an augmentation in the bass under the subject at normal speed, a dominant pedal, a Neapolitan sixth before the final cadence, and a Picardy third at the end. The playback page is a single HTML file with no samples and no libraries. The organ is synthesized in the browser from published measurements of real pipe spectra, down to the chiff when a pipe speaks, and it's tuned in Werckmeister III by default. The model went and found the pipe-spectrum research on its own when I complained the organ sounded fake.

What surprised me was less the music than the method. Instead of writing notes directly into a file and hoping, the model built itself a small toolchain: a notation format, a parser, and a counterpoint checker that flags parallel fifths and octaves, unprepared dissonances, range violations, and voice crossings. Then it composed against the checker, which caught about seventy problems in the first draft, and fixed them in batches.

I should be straight about what this wasn't: a one-shot miracle. It took a handful of rounds, and my ears mattered. The first version of the bass entry had an alto note grinding against it for a full beat. The Neapolitan chord originally arrived straight off the dominant pedal, which is functionally backwards, and it took two rounds of feedback before the model found that diagnosis and fixed the cadence properly. But each time, it located the actual cause in the score and repaired the voice leading rather than papering over it.

A fugue is a good test because the constraints are unforgiving. Every voice has to make sense melodically on its own while every vertical slice makes sense harmonically, and a wrong note isn't a matter of taste. Earlier models couldn't keep that many promises at once. This one mostly can, especially when it's allowed to build its own tools for checking its work.