Voice Think Fast 2.0: the voice that thinks while it talks.
On July 29, SpaceXAI announced Grok Voice Think Fast 2.0, its next-generation speech-to-speech model — the kind of AI that listens to your actual voice and answers out loud, no typing anywhere in the loop. SpaceXAI says it's now the most capable voice model on the market, and the independent benchmark it cites agrees. Here's what changed, what it costs, and why the most interesting user of this model might be the phone line of a small business.
01 The scoreboard
Voice AI has a public leaderboard — the Artificial Analysis Speech-to-Speech Quality Index — and this release moves Grok to the top of it:
| Benchmark | Think Fast 2.0 | Think Fast 1.0 | GPT-Realtime-2.1 | Gemini 3.1 Flash |
|---|---|---|---|---|
| Overall quality index | 82.9% | 75.7% | 79.1% | 69.5% |
| Conversational dynamics | 95.1% | 77.8% | 95.7% | 74.3% |
| Agentic performance (τ-voice) | 56.5% | 52.1% | 45.7% | 37.7% |
| Time to first audio | 0.70s | 1.25s | — | 2.98s |
Two numbers are worth a second look. 0.70 seconds to first audio is roughly the gap in human conversation — the model starts answering about when a person would. And the conversational dynamics jump from 77.8% to 95.1% measures the unglamorous stuff that makes voice AI feel human or robotic: knowing when to talk, when to stop, and how to handle being interrupted.
02 The trick: it reasons while it speaks
Most voice assistants face a trade-off: answer instantly and shallowly, or think first and leave you hanging in silence. The Think Fast line splits the difference — it reasons in parallel with its own speech. While the model is saying its first sentence, it's already working through the harder parts of your request and firing off tool calls in the background. SpaceXAI says 2.0 does this with about 60% fewer reasoning tokens than 1.0, which in practice means lookups and actions usually finish before the model reaches the end of its opening sentence.
It's also been trained — with reinforcement learning against patterns from real human conversations — to talk like a person: shorter sentences, one question at a time, no filler monologues.
03 It out-transcribes the transcription specialists
The part that surprised us: SpaceXAI tested it against dedicated speech-to-text models — tools whose whole job is transcription — across thousands of short phrases in 24 languages. Think Fast 2.0 made 1.5–2× fewer word errors than Deepgram Nova 3 and ElevenLabs Scribe v2, and the gap grew to roughly 10× in noisy conditions: background chatter, phone-line compression, the real world. If you've ever had a voice assistant confidently mishear a name, an address, or a part number, this is the metric that decides whether voice AI is usable for actual work.
04 Who this is actually for
Be clear about what was announced: this is the model behind SpaceXAI's Voice API — the thing developers and businesses use to build talking agents. Two audiences should care:
If you talk to Grok: this is the engine generation behind Grok's voice experiences, and the direction of travel for how the assistant sounds and listens. For what you can do with voice in the apps today — modes, cloning, the voice library — the standing lesson is Voice: modes, cloning, and the voice library.
If you run a business: this is the interesting one. At $0.08 per minute of audio, an AI that answers the phone, understands callers over shop noise, checks your calendar mid-sentence, and books the job costs about $4.80 an hour — and only for the minutes it's actually talking. SpaceXAI's own A/B test runs this model on the Starlink support and sales line, where it reports a significant lift in sales conversion and support containment. If you're a contractor, a shop, or anyone whose phone rings more than you can answer it, "AI receptionist" just got measurably better and stayed cheap.
The migration date matters if you've built anything on it. On August 5, 2026, the grok-voice-latest alias flips from 1.0 to 2.0 automatically. SpaceXAI says existing prompts should just work better — but if you want the old behavior, pin grok-voice-think-fast-1.0 before that date.
05 The honest read
The 60-second version
- Announced July 29, 2026: speech-to-speech model that reasons while it speaks — 0.70s to first audio.
- Top overall score (82.9%) on the Artificial Analysis voice index, ahead of OpenAI and Google.
- Transcription beats dedicated specialist models — by ~10× in noisy, phone-call conditions, per SpaceXAI.
- $0.08/min via the Voice API;
grok-voice-latestauto-upgrades August 5. - Consumer angle: better Grok voice. Business angle: cheap, competent AI phone agents.
Hear the difference yourself
Open Grok on your phone and hold a voice conversation about something with a real answer — "what's the best way to reseal a driveway, and what does it cost?" Interrupt it mid-sentence. Change the subject. Notice how it recovers. That recovery behavior is exactly what the conversational-dynamics benchmark measures — and where this generation moved most.