xAI is swapping the default voice model behind Grok on August 5, 2026. Grok Voice Think Fast 2.0 replaces version 1.0 as the model served under `grok-voice-latest`, and if you're building on the API and haven't pinned a version, your app's voice behavior is about to change whether you asked for it or not.
What's new in 2.0
Think Fast 2.0 is a speech-to-speech model, meaning it doesn't transcribe to text and back, it reasons and responds in audio directly. xAI's own benchmarks show a 1.4x improvement over Think Fast 1.0 in evaluations across thousands of short phrases spanning 24 languages, and a 1.5 to 2.0x improvement relative to Deepgram Nova 3 and ElevenLabs Scribe v2 on the same tests.
The practical upgrades that matter for anyone building a voice agent:
- Better transcription accuracy across languages, not just English
- Faster inference, which shows up as lower latency in conversation
- Improved turn-taking and tool-use reliability, so the model interrupts less and hands off to function calls more cleanly
Pricing
Think Fast 2.0 costs $0.08 per minute of audio, up from $0.05/min on version 1.0. That's a 60% price increase alongside the capability jump.
| Think Fast 1.0 | Think Fast 2.0 | |
|---|---|---|
| Price per minute | $0.05 | $0.08 |
| Price per hour | $3.00 | $4.80 |
| Default as of | Superseded | August 5, 2026 |
If you're running high call volumes, that jump from $3 to $4.80 an hour adds up fast. Budget for it now rather than getting surprised by the invoice in September.
How to keep using the old model
If your app depends on 1.0's specific behavior, or you just want to avoid the cost increase while you test, you need to explicitly pin the model string to `grok-voice-think-fast-1.0` in your API calls. Anyone using the generic `grok-voice-latest` alias gets migrated automatically on August 5, no opt-out.
Where this fits against the competition
xAI is positioning Think Fast 2.0 against Deepgram and ElevenLabs on transcription accuracy specifically, which is a narrower claim than "best voice model overall." If your use case is more about voice quality and emotional range than raw transcription speed, it's still worth testing [ElevenLabs](https://questloops.com/blog/how-to-use-elevenlabs-for-free-in-2026-credits-limits-and-the-catch) side by side rather than assuming xAI's benchmark numbers translate directly to your workload.
The bottom line
This is a solid upgrade if you're already on the Grok stack and need lower latency, better multilingual transcription, or fewer awkward interruptions in your voice agent. It's not automatically a reason to switch stacks if you're happy with what you're running now. Just don't let the automatic default migration catch your production app off guard on August 5.
