StepAudio 3 Realtime
Release date unpublished
StepFun's full-duplex voice model on its Realtime API. It reasons while it speaks and runs tool calls in the background without breaking off the conversation.
- Price per minute
- $0.13/min
- Inputs
- audio, text
- Tool calling
- Yes
- Languages
- 2
- Voices
- —
- Big Bench Audio
- 99.7%
- Access
- Beta
- Licence
- Proprietary
Strengths & weaknesses
Strengths
- Highest Big Bench Audio on AA, 99.7%
- 98.9 on AA Full-Duplex Bench (StepFun technical report)
- Built-in web_search and retrieval tools, run asynchronously
Weaknesses
- Only Chinese and English are fully supported; other languages are in preview
- Still a preview ID during the free trial, with paid billing from 2026-11-09
- No AA Speech to Speech Index score, so its agentic and arena results are unranked
Notable: #1 on AA Big Bench Audio (99.7%); it has no AA S2S Index score. No primary release date: the technical report (arXiv 2609.14005) was first submitted 2026-09-12, and secondary coverage says API availability began 2026-09-15. Model ID: the docs list stepaudio-3-realtime-preview (free trial, to be retired when paid billing starts), and the pricing page lists stepaudio-3-realtime. Report figures: τ-Voice 56.0% macro task success, MMSU 90.6. Open weights not stated, and secondary sources say none. Secondary coverage (remio.ai, aiweekly) reports an 8.83s average time to first audio on AA's benchmark; unverified, so it isn't used.
Links
API model strings
- stepfun
stepaudio-3-realtime - stepfun
stepaudio-3-realtime-preview