Grok Voice Think Fast 2.0
Grok Voice moves from chat to tasks.
On July 29, 2026, xAI announced a major upgrade to its voice model, "Grok Voice Think Fast 2.0." The focus has shifted from how naturally it talks to whether it can actually get things done — bookings, calls, filing.
From "can it talk"
to "can it act"
The competitive axis for voice AI is shifting from conversational fluency to task completion.
Until now, the voice-AI race has mostly been about talking naturally. But xAI's announcement on July 29, 2026 for Grok Voice Think Fast 2.0 is built around a different question: can it finish the task, not just sound good doing it. The context is Voice Agent Builder, the no-code voice-agent platform xAI launched on July 1. That platform gave people an easy way to have voice agents handle bookings, call handling, and paperwork without a human in the loop — and now the "brain" running on top of it, Think Fast 2.0, has been retuned specifically for execution.
The sequencing matters here. First make agents easy to build, then upgrade the model's own ability to actually execute — this is the second wave in voice AI's move from casual conversation partner to operational infrastructure.
Backed by independent benchmarks, too
Not just xAI's own claims — a third-party benchmark ranks it near the top as well.
| Model | S2S Index |
|---|---|
| Grok Voice Think Fast 2.0 | 82.9% |
| OpenAI GPT-Realtime-2.1 | 79.1% |
| Grok Voice Think Fast 1.0 | 75.7% |
| Google Gemini 3.1 Flash | 69.5% |
According to independent AI benchmarking group Artificial Analysis, Think Fast 2.0 ranks #2 overall on the Speech-to-Speech Index and #1 across all models on "Tau Voice for Agentic Performance" — the metric that specifically measures task-completion ability — at 56.5%. It's also the only model among the top five that averages under one second to first audio. TestingCatalog reported the same result as evidence xAI is optimizing for execution over raw conversational polish.
Fewer reasoning tokens,
acting before it finishes talking
Not a smaller model — a model that spends less time "thinking" before it moves.
Think Fast 2.0 cuts reasoning-token usage by roughly 60% versus v1.0 (median relative token use is about 0.4x). That means tool calls — a booking API, a calendar write — often complete before the agent finishes saying its first sentence. On transcription accuracy, xAI says its internal 24-language evaluation shows 1.5–2.0x better word-error-rate than Deepgram Nova 3 and ElevenLabs Scribe v2 — though that's xAI's own internal claim, distinct from the independent Artificial Analysis numbers above, a distinction worth keeping separate (more on this below).
Understand the voice
Parses speech in real time while holding onto context and intent for the next step.
Cut the thinking
Reasoning-token use drops sharply so only the necessary decisions get made, fast.
Act first
Tool calls and API execution start in parallel with speech, returning first audio in 0.70s.
Real work an agent can finish end-to-end
Through Voice Agent Builder, calls, bookings, and paperwork can be completed by voice alone.
Bookings & reservations
Restaurant or clinic bookings get written straight to a calendar mid-conversation, cutting scheduling errors.
First-line call handling
Answers calls, captures the request, and routes or schedules a callback — standing in for human front-desk work.
Filing & data entry
What's said out loud goes straight into a form or spreadsheet, with files sorted automatically after.
The impact, by role
Engineers
If you already have a voice agent running on xAI's platform, the switch to Think Fast 2.0 is close to drop-in. But the grok-voice-latest API alias auto-switches from v1.0 to v2.0 on August 5 — any integration pinned to "latest" will silently change behavior that day, so pin a version now if you need stable behavior.
Business
At $0.08/minute, automated phone and booking workflows tend to undercut the cost of human staffing, especially for high-volume, repetitive interactions.
Product
For task-automation use cases, the #1 ranking on Tau Voice for Agentic Performance is the more relevant signal than general conversational-quality scores.
Talking well isn't the same as finishing the job. Voice built to finish it.
Three things to watch
Pin your version
Before August 5's auto-switch, pin the model version explicitly in any production system that needs stable behavior.
Benchmark your own case
Don't take a single index at face value — run your own booking or inquiry flows through it before deciding.
Watch competitors respond
OpenAI's Presence launched around the same time; how it responds competitively is the next thing to track.
What not to take at face value
The "1.5–2.0x better than Deepgram Nova 3 / ElevenLabs Scribe v2" transcription-accuracy figure is xAI's own internal evaluation, not something independently verified by a third party like Artificial Analysis. The S2S Index and agentic-performance numbers above have third-party backing; the self-reported WER figure doesn't, and that distinction in confidence should be kept separate.
There's also a broader question: as voice agents autonomously carry out real-world actions — bookings, phone calls — without a human checking each step, the operational stakes around detecting mistakes and assigning accountability get heavier. That's a different category of risk than a chat-only assistant ever carried.