Grok Voice Think Fast 2.0: Speech-to-Speech API Guide

Speech-to-speech that reasons while it talks — now with fewer tokens and a hard alias cutover date.

On July 29, 2026, xAI (SpaceXAI) announced Grok Voice Think Fast 2.0: better speech reasoning, 1.5–2× transcription WER vs dedicated STT on their 24-language eval, ~60% fewer reasoning tokens than 1.0, and snappier tool timing. grok-voice-latest flips to 2.0 on August 5, 2026. List price: $0.08/min of audio.

This is the API companion to our Voice Agent Builder coverage — raw WebSocket / Realtime-compatible surface for assistants, phone agents, and interactive voice, not just the no-code console.

TL;DR

What people are asking

“Is this only a model bump or a new API?”

Both. The model is Think Fast 2.0. The Speech to Speech docs describe the production WebSocket surface: session params, VAD, codecs, binary audio transport, tools (search, collections, MCP, functions), force_message, session resumption, pronunciation replace, and OpenAI Realtime migration notes. Docs last updated July 27, 2026 in the paste you shared — treat the announcement (Jul 29) as the model story and the docs as the integration contract.

“Why did price go to $0.08/min?”

Earlier Voice Agent Builder coverage cited $0.05/min. Think Fast 2.0’s announcement lists $0.08/min. Budget on the current pricing page, not July 1 blog posts. Telephony add-ons (if any) still stack separately in the Builder path.

“Will my prompts break on Aug 5?”

xAI claims most prompts improve with no edits, and Starlink A/B (+1 888 GO STARLINK) saw higher sales conversion and support containment. Migration guidance for 2.0 still says: shorten system prompts, drop GPT-era workaround instructions, and remember reasoning.effort defaults to "high" — set "none" if you want less parallel thinking.

Production rule: pin a versioned ID (grok-voice-think-fast-2.0 or 1.0). Let grok-voice-latest float only in staging.

“How does this compare to OpenAI Realtime / GPT Live?”

Independent of marketing: run your call scripts, noise profiles, and tool schemas. AA and τ-voice are directional.

Artificial Analysis numbers (as published)

From xAI’s announcement, source Artificial Analysis:

Transcription: xAI claims 1.5–2× better WER vs Deepgram Nova 3 / ElevenLabs Scribe v2 on thousands of short phrases across 24 languages, ~1.4× vs TF 1.0, and up to ~10× better in noisy telephony conditions. Those are vendor evals — reproduce on your recordings.

Quick start (WebSocket)

import asyncio, json, os, websockets

MODEL = "grok-voice-think-fast-2.0"  # or grok-voice-latest after Aug 5
URL = f"wss://api.x.ai/v1/realtime?model={MODEL}"

async def main():
    async with websockets.connect(
        URL,
        additional_headers={"Authorization": f"Bearer {os.environ['XAI_API_KEY']}"},
    ) as ws:
        await ws.send(json.dumps({
            "type": "session.update",
            "session": {
                "voice": "eve",
                "instructions": "You are a helpful assistant. Keep turns short.",
                "reasoning": {"effort": "high"},  # or "none"
                "turn_detection": {"type": "server_vad"},
            },
        }))
        await ws.send(json.dumps({
            "type": "conversation.item.create",
            "item": {
                "type": "message",
                "role": "user",
                "content": [{"type": "input_text", "text": "Hello!"}],
            },
        }))
        await ws.send(json.dumps({"type": "response.create"}))
        async for raw in ws:
            print(json.loads(raw).get("type"))

asyncio.run(main())

Auth: ephemeral tokens for browsers/mobile; API key server-side only. Tester apps in docs: iOS, WebSocket web, WebRTC, Twilio telephony.

Session knobs that matter in production

Codecs: PCM (8–48 kHz), μ-law / A-law (8 kHz), Opus (24 kHz). Default PCM 24 kHz.

Tools: search, collections, MCP, functions

Configure on session.update:

• file_search — collections / vector stores
• web_search / x_search — server-side, optional domain/handle filters
• mcp — remote MCP over streaming HTTP/SSE; xAI executes server-side
• function — client executes; handle response.function_call_arguments.done

Parallel tools: collect all function_call_output items before response.create.
Audio overlap trap: wait for current playback (or near end) before response.create after tools, or the next utterance stacks on the previous.

Migrating from OpenAI Realtime

1. base_url="https://api.x.ai/v1" + XAI_API_KEY (or WebSocket host wss://api.x.ai/v1/realtime).
2. Model → grok-voice-think-fast-2.0 / grok-voice-latest.
3. Voice → Grok roster (eve, ara, rex, …) or custom voice ID.
4. Simplify prompts; drop GPT workaround strings.
5. Check incompatibilities: e.g. transcription event is …updated (cumulative) not OpenAI’s …delta when using grok-transcribe; several output_audio_buffer.* / conversation.item.retrieve events unsupported.
6. Adopt xAI extensions only where needed (force_message, resumption, replace).

Compare stacks with GPT-Realtime-2, GPT Live, and Claude Voice.

Builder checklist before Aug 5

1. Pin grok-voice-think-fast-1.0 in prod if you are not ready for 2.0.
2. A/B 2.0 on staging with real telephony noise and your tool schemas.
3. Re-cost at $0.08/min vs prior $0.05 assumptions.
4. Enable binary transport if you control both ends.
5. Add keyterms for product names; replace for pronunciation.
6. Implement tool → wait-for-playback → response.create ordering.
7. Prefer ephemeral tokens for any client that is not your server.
8. If you use Voice Agent Builder, confirm which runtime model the console pins after Aug 5.

Cutover plan
□ Staging pinned to 2.0 since this week
□ Prod still on 1.0 OR explicitly on 2.0 (no surprise latest)
□ Cost model updated to $0.08/min
□ Prompt slimmed; reasoning.effort decided
□ Playback-gated tool continuation tested

Telephony and enterprise path

Docs position Speech to Speech for PSTN / contact-center / PBX via SIP, WebSocket, or LiveKit, with native G.711 μ-law / A-law so you are not forced through PCM transcoding on every hop. Pair that with force_message for recorded-call disclosures and IVR lines that must be verbatim TTS — without waiting on the model to invent the script. Session resumption matters when mobile clients drop WebSockets: capture conversation.created.conversation.id, reconnect with ?conversation_id=, and re-send resumption.enabled: true so transcripts and tool outputs replay (history expires after 30 minutes idle).

If you already run agents in the Voice Agent Builder console, treat Think Fast 2.0 as the runtime brain under that UI: confirm which model ID the console pins after August 5, and whether Builder metering still shows the older $0.05/min line item or has moved to the announcement’s $0.08/min Think Fast 2.0 rate. Do not assume Builder and raw Realtime share one invoice story until finance pulls a fresh usage export.

Languages: English plus regional Arabic, Spanish, Portuguese, and many others (zh, ja, ko, hi, …). Use language_hint when callers stay in one locale; leave auto-detect for mixed queues. Add keyterms for SKUs and proper nouns that ASR otherwise mangles.

Honest limitations

• AA / WER / τ-voice figures are vendor-published (Artificial Analysis + xAI evals).
• Price and alias dates can move — re-read pricing and docs.
• Conversational Dynamics still lists GPT-Realtime-2.1 slightly ahead (95.7% vs 95.1%).
• OpenAI Realtime “compatible” ≠ identical event surface.
• Builder $0.05/min vs API $0.08/min can confuse finance — clarify which product you buy.
• 20+ listed languages; others “capable with varying accuracy.”
• Parallel reasoning still costs tokens even at 0.4× of 1.0 — set reasoning.effort: "none" for scripted IVR if latency and cost dominate.

Closing

Think Fast 2.0 is the voice model to evaluate if you already bet on Grok’s speech-to-speech path — or if you want a Realtime-shaped WebSocket with $/min metering, MCP, and telephony-friendly codecs. Pin before August 5, re-run your call scripts, and treat Starlink conversion anecdotes as a hint, not your KPI. Ship binary transport and keyterms only after you have a golden-path recording of your noisiest line; model upgrades do not fix a bad audio pipeline.

Follow @explainx_ai for voice-stack follow-ups.

Related on explainx.ai

• xAI Grok Voice Agent Builder — no-code / $0.05 era
• OpenAI GPT-Realtime-2 Voice Models API
• GPT Live — ChatGPT voice
• Claude Voice mode
• OpenAI Presence — enterprise voice
• Grok 4.5 public launch
• SpaceXAI Grok Build Mode
• Miso One open voice model
• Voicebox open-source studio

Sources

• Introducing Grok Voice Think Fast 2.0
• xAI / SpaceXAI Speech to Speech docs (Voice → Speech to Speech)
• Artificial Analysis — Speech to Speech
• xAI pricing

Model IDs, alias cutover (August 5, 2026), benchmark tables, and $0.08/min pricing as published by xAI on July 29, 2026. Re-verify docs and pricing before production cutovers.