Last updated: October 29, 2026
Grok Voice 3.0 is xAI’s newest speech-to-speech model, announced October 14, 2026. Developers call it as grok-voice-3.0 (or the grok-voice-latest alias) over a WebSocket Realtime API. It costs $0.08 per minute, the same as Grok Voice 2.0, and needs no code changes beyond the model name.
xAI, which now signs its documentation as SpaceXSI, says the model has latency under one second and powers voice conversations in Tesla vehicles and Grok Bot’s voice mode (xAI announcement).
Contents
- What’s new
- Grok Voice 3.0 vs 2.0 and rivals
- How to start
- Key session settings
- Tools and phone calls
- Pricing and limits
- Limitations
- Frequently asked questions
What’s new in Grok Voice 3.0
The xAI release notes list three October changes for voice developers: grok-voice-3.0 is available in Speech to Speech, grok-voice-latest now points to it, and grok-voice-2.0 is deprecated. The API itself is unchanged.
The practical changes are in model quality. xAI’s migration notes say the model is significantly more capable, so system prompts should be much shorter, and workaround instructions written for older models can be removed. Reasoning is enabled by default at high effort.
xAI also says Starlink call centers use Grok Voice to answer 20,000 calls a day, with roughly 70% resolved without a human. These are company figures, not independent results.
Grok Voice 3.0 vs Grok Voice 2.0 and rivals
xAI published these results from Artificial Analysis testing dated October 13, 2026 (benchmark table). Higher is better on every row.
| Benchmark | Grok Voice 3.0 | Grok Voice 2.0 | GPT-Live-1 (Astra, Medium) | Gemini 3.8 Live |
|---|---|---|---|---|
| AA Speech-to-Speech Index | 86.9 | 81.3 | 83.2 | 82.6 |
| Big Bench Audio (speech reasoning) | 98.5% | 97.2% | 90.1% | 97.7% |
| Speech Agent Arena Elo | 1070 | 1011 | 1048 | 990 |
| Task success rate | 94.8% | 94.6% | 87.4% | 89.1% |
| Full Duplex Bench | 97.0% | 95.1% | 94.9% | 91.9% |
On the τ-Voice support-call test, xAI reports 71.6%, comparable with GPT-Live-1 (Astra) at 40% lower latency. ServiceNow’s EVA-Bench gave it the top accuracy score of 74%, according to the announcement. Price and API shape match Grok Voice 2.0, so the upgrade is a quality change only.
If you are comparing text models from the same week, see our GPT-6.1 Terra power-user guide and our Gemini 4 Argon guide.
How to start with Grok Voice 3.0
- Create an API key in the xAI console. Keep it on your server.
- Open a WebSocket to
wss://api.x.ai/v1/realtime?model=grok-voice-3.0with anAuthorization: Bearerheader. Pin the versioned name in production; usegrok-voice-latestonly if you want automatic upgrades. - Send
session.updatewithinstructions, avoicesuch aseve, andturn_detectionset toserver_vad. - Stream audio with
input_audio_buffer.appendand play eachresponse.output_audio.deltaas it arrives. - For browsers and mobile apps, use an ephemeral token passed in the WebSocket protocol instead of your API key.
A minimal session configuration from the Speech to Speech docs looks like this:
{"type": "session.update",
"session": {
"voice": "eve",
"instructions": "You are a helpful support agent.",
"turn_detection": {"type": "server_vad"},
"reasoning": {"effort": "high"}
}}Coming from the OpenAI Realtime API? xAI says most OpenAI SDKs work after you change the base URL to https://api.x.ai/v1, swap the API key and choose a Grok voice model.
Key session settings for power users
reasoning.effort:high(default) ornone. Usenonefor short, scripted flows where speed matters.- Audio formats: PCM (default 24 kHz; 8 kHz to 48 kHz supported), G.711 μ-law and A-law at 8 kHz for telephony, and Opus at 24 kHz. Set
transporttobinaryto send raw frames instead of base64 JSON. - Keyterms: up to 100 terms, each up to 50 characters, to bias transcription toward product names.
- Language hint: a BCP-47 code such as
jaores-MX; Spanish and Portuguese need a regional variant. - Pronunciation
replacemap: changes only the spoken audio, not the transcript. - Output speed: 0.7 to 1.5, default 1.0.
- Session resumption: set
resumption.enabledand reconnect withconversation_id; history is dropped after 30 minutes of inactivity.
Two xAI-specific features help with compliance. A force_message item speaks fixed text verbatim, such as a recording disclosure, and can be made non-interruptible. Per-response instructions on response.create override the system prompt for one turn, which is useful for injecting caller or CRM data.
Tools and phone calls
Grok Voice 3.0 sessions can use web_search, x_search, file_search over your document collections, remote mcp servers and custom function tools. When several functions are called at once, send every function_call_output before a single response.create, and wait for current audio to finish playing to avoid overlapping speech.
For phone agents, the docs describe SIP and LiveKit telephony integrations, and G.711 audio formats work directly with phone networks. The model supports more than 20 languages with automatic language detection.
For background on outages that can affect voice products, see our explainer on whether ChatGPT, Claude and Grok are down. For coding agents rather than voice, compare our GitHub Copilot Fusion guide.
Grok Voice 3.0 pricing and limits
Prices below come from the xAI pricing page, and session limits from the rate limits page.
| Item | Official figure |
|---|---|
| Speech to Speech (grok-voice-3.0) | $0.08 per minute ($4.80 per hour) plus $0.004 per text input |
| Web Search or X Search tool | $5 per 1,000 calls (X Search: per 1,000 posts) |
| Speech to Text (separate API) | $0.10 per hour REST, $0.20 per hour streaming |
| Text to Speech (separate API) | $15.00 per 1 million characters |
| Concurrent sessions | 10 (Tier 0), 20 (Tier 1, $50 spend), 50 (Tier 2, $250), 100 (Tier 3, $1,000), 200 (Tier 4, $5,000) |
Speech to Speech is limited by concurrent sessions alone, not tokens. Tiers are based on total spend, and once reached they never downgrade. Enterprise limits are available on request.
Limitations to know
- Benchmark figures are published by xAI from Artificial Analysis and ServiceNow tests, not by an independent review of your use case.
- The
grok-voice-latestalias changes models without notice, so production apps should pingrok-voice-3.0. - Some OpenAI Realtime events are not supported, including
conversation.item.retrieveandrate_limits.updated. - Without session resumption, conversation history is lost when the WebSocket closes.
- Grok Voice 2.0 is deprecated, so plan to move pinned sessions now.
Frequently asked questions
What is Grok Voice 3.0?
Grok Voice 3.0 is xAI’s speech-to-speech model for real-time voice agents, announced on October 14, 2026. Its API model name is grok-voice-3.0, and it runs over the Speech to Speech API at wss://api.x.ai/v1/realtime. It also powers Grok Bot’s voice mode and voice conversations in Tesla vehicles.
How much does Grok Voice 3.0 cost?
The Speech to Speech API charges $0.08 per minute ($4.80 per hour) for grok-voice-3.0, the same price as Grok Voice 2.0, plus $0.004 per text input. Server-side tools such as web search are billed separately, for example $5 per 1,000 web search calls.
How do I upgrade from Grok Voice 2.0 to 3.0?
If your app connects with grok-voice-latest, it already runs on Grok Voice 3.0. If you pinned grok-voice-2.0, change the model query parameter to grok-voice-3.0. The endpoint, session parameters, events and price are unchanged, so no other code changes are required.
How many concurrent Grok Voice 3.0 sessions can I run?
Speech to Speech is limited by concurrent sessions. The xAI rate limits page lists 10 sessions at Tier 0, 20 at Tier 1, 50 at Tier 2, 100 at Tier 3 and 200 at Tier 4. Tiers are based on total spend, starting at $50 for Tier 1, and never downgrade.
Does Grok Voice 3.0 work with OpenAI Realtime API code?
Yes, mostly. xAI says the Speech to Speech API is compatible with the OpenAI Realtime API: change the base URL to api.x.ai, use an xAI API key and pick a Grok voice model. A few events are named differently or not emitted, so check the compatibility table.
Can I turn off reasoning in Grok Voice 3.0?
Yes. Reasoning is on by default with reasoning.effort set to high, which helps with multi-step instructions and ambiguous requests. For simple tasks where speed matters more, set reasoning.effort to none in your session.update event.
Sources
- xAI: Introducing Grok Voice 3.0
- xAI Docs: Speech to Speech
- xAI Docs: Release Notes
- xAI Docs: API Pricing
- xAI Docs: Rate Limits
Image: Grok logo by xAI, public domain, via Wikimedia Commons. Placed on a white background and resized.
