Grok's voice alias flips August 5 and your rate goes up 60%
xAI moves grok-voice-latest to Think Fast 2.0 on August 5. Same code, $0.05 to $0.08 per audio minute. Pin your model version or price the change now.
If your voice agent calls grok-voice-latest, it gets a different model tomorrow. On August 5, xAI moves that alias from Grok Voice Think Fast 1.0 to 2.0. You don't have to do anything — which is the problem, because model aliases are how a pricing change arrives without a purchase order.
What actually happened
xAI launched Grok Voice Think Fast 2.0 on July 29. It's a real upgrade: time to first audio drops from 1.25 seconds to 0.70, and it scores 82.9% on Artificial Analysis' speech-to-speech index against 75.7% for 1.0, ahead of GPT-Realtime-2.1 at 79.1% and Gemini 3.1 Flash at 69.5%.
It also costs more. xAI's own model docs list Think Fast 1.0 at $0.05 per audio minute and 2.0 at $0.08 — a 60% increase per minute of conversation. Developers who want to stay on 1.0 have to pin grok-voice-think-fast-1.0 before the switch. Everyone else migrates by default.
Why it matters for your business
Run the arithmetic on your own call volume, because "3 cents" is not the unit that matters. A front-desk agent handling 400 calls a month at four minutes each is 1,600 audio minutes: $80 becomes $128. Ten thousand minutes goes from $500 to $800. That's a line item that moves without anyone approving it, and it shows up on a statement three weeks after the fact.
The latency win is probably worth the money — 550ms off time-to-first-audio is the difference between a caller thinking the line dropped and thinking they're talking to someone. But that's a decision you should make, not a default you inherit. Two rules that generalize past xAI: never ship -latest to production, and keep the model identifier in config rather than compiled into your call handler, so switching is an environment variable instead of a deploy. Then run both versions against twenty recorded calls from your actual business, with your actual accents and background noise, and compare interruption handling and transcription accuracy before you commit. Vendor benchmarks are measured on someone else's audio.
We build voice agents where the model is a config value and the swap takes five minutes.
Key takeaways
- On August 5, 2026, xAI's
grok-voice-latestalias moves from Think Fast 1.0 to 2.0 automatically - Per xAI's docs, that's $0.05 to $0.08 per audio minute — a 60% rate increase with no action on your part
- 2.0 is genuinely faster (0.70s vs 1.25s to first audio) and scores 82.9% vs 75.7% on Artificial Analysis' speech-to-speech index
- Pin explicit model versions in production, keep the identifier in config, and A/B the upgrade against your own recorded calls before accepting it
Running a voice agent you didn't build? We pin the model, put the identifier in config, and test upgrades against your real call recordings before they hit customers. See the voice systems we've shipped or book a build call.
Sources: xAI, xAI model docs.
- #voice-ai
- #model-pinning
- #xai
- #api-pricing
- #vendor-risk
Tommy Rush — Founder, Rush Commerce
Operator turned builder. 15+ years running operations — now shipping the systems businesses run on. More
Get The Rush Report weekly — one email, zero fluff.
Keep reading
Valar Atomics raises $1B: your AI bill is a power bill
Sequoia led a $1B round at a $6B valuation for factory-built nuclear reactors aimed at AI data centers. Why compute scarcity is now an electricity problem.
Read itRunware's shipping-container data center for AI inference
Runware unveiled a portable AI inference pod: 1,200 GPUs and 1MW in a 20-foot container, built in weeks. Why where your inference runs sets your latency and price.
Read it