Stability AI releases Stable Audio 2.5 for enterprises

Background

Audio generation models take a text prompt and produce music, sound effects or other audio. For an individual user that is a novelty; for organisations that produce advertising, games, video and interactive media it is a possible supply of sound at volume, because a single campaign can call for dozens of short cues, each of which has to be cleared for commercial use. Two questions tend to decide whether such a tool enters a real production pipeline: whether the training data is licensed in a way that permits commercial output, and whether generation is fast enough to sit inside an existing workflow rather than beside it.

Release

On 10 September 2025 Stability AI released Stable Audio 2.5 ^[1]. The company described it as the first audio model built for enterprise sound production at scale ^[1], and said it was trained on fully licensed data ^[1]. Both points are Stability's own framing rather than an independent assessment, but together they show where the model was placed: not as a general creative toy, but as a tool for organisations that need to know the rights position before they ship anything.

Speed and training

Stability reported an inference speed of less than two seconds on a GPU for tracks up to three minutes ^[3]. Inference is the step in which a trained model actually produces output, as opposed to the training that built it; a couple of seconds per track matters when a team wants to generate many variations and pick one, because waiting turns a creative loop into a batch job. The figure is for a GPU, the hardware accelerator commonly used to run such models.

The model was post-trained with the Adversarial Relativistic-Contrastive (ARC) method developed by the Stable Audio research team ^[3]. Post-training is the stage that follows a model's main training run, in which an already trained system is adjusted further; naming a specific method signals that the team treats this stage as part of its own research rather than as routine tuning.

How it is offered

Stable Audio 2.5 was made available on StableAudio.com, through the Stability AI API, on partner platforms and on-premises under enterprise licences ^[1]. The partner platforms named were fal, Replicate and ComfyUI ^[3]. These are different routes to the same model: a website for direct use, an API for developers who want to build audio into their own software, third-party platforms where many models are available side by side, and on-premises deployment, in which the model runs on the customer's own machines. That last option matters to studios and brands that cannot send material to an outside service.

Brand sound identities

Stability partnered with Amp, WPP's sound branding agency, on brand sound identities ^[2]. A sound identity is the audio counterpart of a logo: a short, repeatable sound that accompanies a brand across advertising and other material. Stability cited Ipsos research that custom audio can make a brand eight times more memorable while only 6% of creative work uses a sound identity ^[3]. The cited figures are the commercial argument for the product — a large gap between the reported effect of custom sound and how rarely it is used.

What it signals

Stable Audio 2.5 is a small release that says something about direction. Licensing is presented as a feature rather than a footnote, which is what an enterprise buyer asks about first with generated media. Generation speed is quoted as a production property rather than a benchmark score. And the model is distributed widely, from a consumer-facing site to on-premises installations, so that the same system can be tried by one person and bought by a studio. The claims about scale, rights and memorability come from Stability and its cited research, and the case for the model ultimately rests on whether sound teams adopt it for work they ship.