# What’s new at HeyGen: September 2026 By Steven Van · 2026-10-01 Developers get MIT-licensed avatar demos, a 168-brief motion-graphics benchmark and API voice cloning from recorded speech. Developers can now build with three September releases from [HeyGen](): an open-source real-time avatar stack, a motion-graphics benchmark and Professional Voice Clone through the API. All three are live, according to the [October 1 update](). ## Real-time avatars with on-screen visuals Released on September 10, the reference integration connects OpenAI’s GPT-Live-1 speech-to-speech model to a HeyGen LiveAvatar, with tool calls displayed as animated HyperFrames overlays. It includes working language tutor, poker coach and customer support demos. Developers can install dependencies, run setup and start the demo with three commands, then change two markdown files to adapt the persona. A second demo generates motion graphics during a conversation. A background Claude Code session writes a HyperFrames composition that appears over the avatar’s video. Generation takes 30 seconds to two minutes, while the avatar acknowledges the request and keeps talking. Both repositories are MIT licensed. A LiveAvatar API key is required, plus an OpenAI key with GPT-Live access for the first demo. HeyGen describes them as starters and says developers should add authentication before exposing them publicly. ## A benchmark for generated motion graphics Released with Kaggle on September 18, Code2Video evaluates how well models turn creative briefs into motion-graphics code rendered by HyperFrames. Its **168 briefs** cover parts of launch videos, each paired with a human-made reference. Scores assess engagement, prompt intent, composition, timing and craft. HeyGen reports that its purpose-built judge agrees with human preferences 82% of the time and returns a verdict in under a second. The open leaderboard uses Elo rankings, with runs and raw traces on Kaggle. Early results show that following the brief is nearly solved, while craft remains weaker, particularly motion. Running custom models against the same judge is described as coming next. ## Voice cloning through the API Professional Voice Clone trains a dedicated voice adapter from at least 20 minutes of one speaker’s recordings. The workflow is to upload one to 10 recordings through the Assets API, create the voice, poll until it is active, then request a finished WAV or streamed audio. Creators can update or localize narration without another recording session, and training teams can reuse an instructor’s voice across courses. Retraining with the existing voice ID preserves its ID, name, language and slot. If retraining fails, the previous voice stays active. The docs also include a prompt for Claude Code, Codex or Cursor to guide users through cloning their own voice. Each voice uses a purchased API slot, which includes five trainings per billing period, counting the first training. Synthesis costs **0.6 API credits per generated minute**. Voices belong to the workspace associated with the API key that created them, and deleting a voice frees its slot. [![]()]() ![HeyGen]() HeyGen Talking-head videos from AI avatars — pick from hundreds or build your own from a clip, script in 40+ languages, and translate with lip sync intact. [View HeyGen →]() [Original source]() [Read What’s new at HeyGen: September 2026 on Creators Toolbox]() --- Canonical source: https://creatorstoolbox.com/blog/heygen-what-s-new-at-heygen-september-2026