Stability AI Ships Stable Audio 2.5, Targets Enterprise Sound Branding

Stability AI has released Stable Audio 2.5, its first audio generation model built explicitly for enterprise use, promising faster generation, better musical structure, and new editing tools aimed at professional creative teams.

The model can produce a three-minute track in under two seconds on a GPU, thanks to a post-training technique the company calls Adversarial Relativistic-Contrastive (ARC). Beyond speed, the output is musically structured — generating full compositions with intros, developments, and outros — and responds more precisely to mood and genre prompts like "uplifting" or "lush synthesizers."

New to this version is audio inpainting. Users can upload their own audio, select a starting point, and let the model generate the rest of the track in context. Stability AI's terms require that any uploaded audio be free of copyrighted material, and the company uses content recognition tools to enforce that.

The enterprise play is clear. Stability AI is partnering with amp, a sound branding agency within WPP's Landor Group, to co-develop custom audio solutions for major brands. Enterprise clients can also have Stable Audio models fine-tuned on their own sound libraries, embedding signature brand audio directly into generative workflows. The technology will be available to WPP's global client base through the WPP Open platform.

For builders and developers, the model is accessible now through StableAudio.com, the Stability AI API, and partner platforms including fal, Replicate, and ComfyUI. Enterprise licensing with implementation support and customization is also available. Like previous Stable Audio models, 2.5 is trained on a fully licensed dataset and marketed as commercially safe.