# Stability AI introduced Stable Audio 2.5 with ARC post-training and inpainting Stability AI introduced Stable Audio 2.5 on September 10, 2025, combining a stated sub-two-second GPU claim with three-minute tracks, inpainting and enterprise deployment options. Canonical: https://rangoon.ai/insights/archive-stability-audio-2-5-september-2025/ Author: Rangoon Editorial (https://rangoon.ai/insights/editorial/) Published: 2026-10-24 Updated: 2026-10-24 Event date: 2025-09-10 Tags: [Stability AI](https://rangoon.ai/insights/tags/stability-ai/), [audio AI](https://rangoon.ai/insights/tags/audio-ai/) Source note: Historical reporting: Source 0 is the dated Stable Audio 2.5 release. Source 1 is the associated ARC post-training paper, used for method context and not for the September event date. Image: Stable Audio 2.5 generation and inpainting path. Original Rangoon explanatory diagram: Source boundary: dated primary release reporting; original editorial analysis. No vendor interface, photograph, measurement result, or endorsement implied. ## Key takeaways - Stability AI dated Stable Audio 2.5 to 2025-09-10 and named StableAudio.com, API, partner and enterprise routes. - The announcement described text-to-audio, audio-to-audio and inpainting, including tracks up to three minutes and a stated sub-two-second GPU figure. - The technical paper describes ARC as adversarial post-training for diffusion or flow models rather than a blanket speed guarantee. ## The release connected composition length to a speed claim [source 1](https://stability.ai/news-updates/stability-ai-introduces-stable-audio-25-the-first-audio-model-built-for-enterprise-sound-production-at-scale) Stability AI announced Stable Audio 2.5 on September 10, 2025 as an enterprise-oriented audio-generation model. Its central performance statement was not merely that generation was fast: the company said the model could make tracks up to three minutes long within seconds and reported inference below two seconds on a GPU. It also positioned the model for multi-part musical structure, naming intros, development and outros, and said prompt adherence had improved for mood descriptors and musical language. Those statements are specific to the supplier's release and its stated conditions. They do not tell a reader what GPU was used for every request, how an arbitrary prompt will sound, or whether a whole production will take seconds once uploads, approvals and editing are included. The dated announcement establishes what Stability AI chose to claim for this release; it is not a broadly replicated latency study. ## ARC is a post-training method, not a synonym for the product [source 1](https://stability.ai/news-updates/stability-ai-introduces-stable-audio-25-the-first-audio-model-built-for-enterprise-sound-production-at-scale) [source 2](https://arxiv.org/abs/2505.08175) The release attributes the speed work to Adversarial Relativistic-Contrastive, or ARC, post-training. The related paper describes ARC as an adversarial acceleration procedure for diffusion and flow models that does not use distillation. Its authors say the method combines a relativistic adversarial formulation with a contrastive discriminator objective intended to improve prompt adherence. In the paper's Stable Audio Open experiment, the authors report roughly 12 seconds of 44.1 kHz stereo audio in about 75 milliseconds on an H100 and roughly seven seconds on a mobile edge device. Those figures belong to that research configuration, not automatically to the three-minute Stable Audio 2.5 product claim. Keeping those records separate prevents an easy but misleading move: treating a short research sample, a particular accelerator and a later commercial interface as the same measurement. ## Inpainting made the input audio part of the workflow [source 1](https://stability.ai/news-updates/stability-ai-introduces-stable-audio-25-the-first-audio-model-built-for-enterprise-sound-production-at-scale) Stable Audio 2.5 also added an editing shape that differs from generating a track from text alone. In its audio-inpainting workflow, a user can supply audio, choose a start point and ask the model to generate from the surrounding context. The launch says this sits beside text-to-audio and audio-to-audio work. That gives a sound team a possible way to continue or alter a passage rather than regenerate an entire piece. It also brings the uploaded material into scope. Stability's release says its terms require uploads to be free of copyrighted material and that it uses content recognition to maintain compliance and prevent infringement. The announcement describes the supplier's policy and system; it does not transfer rights in a user's source audio or decide a disputed copyright question. ## Distribution channels did not erase deployment differences [source 1](https://stability.ai/news-updates/stability-ai-introduces-stable-audio-25-the-first-audio-model-built-for-enterprise-sound-production-at-scale) [source 2](https://arxiv.org/abs/2505.08175) At launch, Stability AI listed several access routes: StableAudio.com, its API, partner platforms including fal, Replicate and ComfyUI, and an enterprise-license path for on-premises deployment. It also said organizations could discuss fine-tuning models on an organization's sound library. Those routes matter because a browser product, an API integration, a partner service and a customized deployment can carry different account controls, terms and implementation work. The release's claim that the model was trained on a fully licensed dataset is a supplier statement about training data; it is not a conclusion about every generated output, client brief or uploaded file. The durable news fact is a September 10 launch with named editing and access surfaces. The dated record does not prove a sound is exclusive to a brand, clear a composition for a campaign, or guarantee the stated latency on another machine or provider. ## Sources 1. [Stability AI: Stable Audio 2.5 release](https://stability.ai/news-updates/stability-ai-introduces-stable-audio-25-the-first-audio-model-built-for-enterprise-sound-production-at-scale) 2. [Stable Audio research: ARC post-training paper](https://arxiv.org/abs/2505.08175) ## Related reading - [Stability AI and Arm released Stable Audio Open Small](https://rangoon.ai/insights/archive-stability-audio-open-small-may-2025/) - [Suno released v4.5 with longer songs and new prompt controls](https://rangoon.ai/insights/archive-suno-v4-5-may-2025/) - [Stability AI released Stable Virtual Camera as a research preview](https://rangoon.ai/insights/archive-stability-virtual-camera-march-2025/) ## More information - [Documentation index](https://rangoon.ai/llms.txt) - [AI access and policies](https://rangoon.ai/ai/) - [Source repository](https://github.com/hypler-dev/rangoon)