StepAudio 3 Chat Preview is live on the model catalog at $0 per token. The id customers send is stepaudio-3-chat-preview:free — free-lane pricing, no waiting list, no card. Every account with a key can call it now.
It is StepFun’s audio-chat preview, and it is a different kind of model from the rest of the free lane. The others are text LLMs. This one is an audio-chat model that reasons: it reads your message, thinks through the reply (its responses carry a reasoning trace), and answers on the ordinary chat completions endpoint. You send text in and get text back, exactly like any other model on the gateway — so nothing about your integration changes.
What it is, and what it is not
Two clarifications, because the name causes confusion.
It is not a text-to-speech engine. StepFun ships separate TTS models (stepaudio-2.5-tts, stepaudio-3-tts). Those convert a script into audio. StepAudio 3 Chat is the voice-chat tier — an LLM that reasons and replies, built for conversational use. For batch transcription StepFun has the *-asr models; for speech-from-a-script, the *-tts models. This one is the chat model.
It is a preview. StepFun labels it preview and publishes no context window or max output for it. We left those blank in the model catalog rather than guess. Expect StepFun to move the underlying snapshot without notice — that is what “preview” means. Pin the exact id and re-test after any provider announcement.
How to use it
It speaks the OpenAI chat protocol, so the same client you already use works. Point it at the gateway and set the model:
curl https://api.oneaiapikey.com/v1/chat/completions \
-H "Authorization: Bearer $ONEAI_API_KEY" \
-H "Content-Type: application/json" \
-d '{
"model": "stepaudio-3-chat-preview:free",
"messages": [{"role": "user", "content": "Summarise this in one line."}]
}'Use it anywhere a chat model fits: quick Chat sessions, a summariser, a reasoning step inside an agent. At $0 it is free-lane pricing, so it counts against the free-lane daily request allowance rather than your credit. That allowance refills at 00:00 UTC (05:30 IST); the rate limits docs cover the caps.
What we checked before it went live
We probe a model on its own provider before putting it on the free lane. Checked 2026-11-12 and again 2026-11-13 on StepFun’s upstream: a real request returned a completion with a reasoning trace, and four rapid calls in a row all succeeded. The text flagship from the same provider (step-5-preview) is out of quota upstream, so this audio-chat model is the one being added. On our side it is wired into the gateway at ratio 0, so every call bills $0.
For live health, watch the status board, and the changelog for the next additions. Also new this week: a faster free lane for Qwen3.8 27B with 1,000 requests a day. See every $0 model on the free lane.