GROK VOICE THINK FAST 2.0 — CONFIRMED SPECS (JULY 29, 2026)
● Pricing: $0.08/min of audio (up from $0.05/min for v1.0)
● First audio response: 0.70 seconds
● Transcription vs v1.0: 1.4× better overall
● Transcription vs competitors: 1.5-2× better than leading transcription models
● Noisy environments: ~10× better transcription in noisy settings
● Languages: Tested across 24 languages
● Token efficiency: 60% fewer reasoning tokens than v1.0
● Tool calls: Usually begin before the first sentence ends
● Conversational style: Shorter sentences, asks one question at a time, no unnecessary filler
● Migration deadline: August 5 — grok-voice-latest auto-upgrades to 2.0
● Access: xAI API + Voice Agent Builder (no-code)
● Tested by SpaceX: Starlink sales conversion and support containment — both improved
What Changed From v1.0
According to SpaceXAI's official announcement, Grok Voice Think Fast 2.0 improves on three fronts simultaneously: intelligence (better reasoning through complex workflows), transcription accuracy, and conversational naturalness. The predecessor, Think Fast 1.0, launched in April 2026 at $0.05/min and scored 67.3% on Tau-Voice Bench — beating Gemini 3.1 Flash Live and GPT Realtime at 66% and 36% respectively on airline itinerary change tasks. Version 2.0 represents the next generation, not a minor update.
The 60% reduction in reasoning tokens is the most operationally significant number. In a voice agent context, fewer reasoning tokens means lower per-call cost and faster response within the conversation turn. As Basenor's developer breakdown notes, the architecture runs as a single speech-to-speech model — not a stitched pipeline of separate speech-to-text, language model, and text-to-speech components. That's why the latency is 0.70 seconds for the first audio response: there are no pipeline hops adding delay between hearing the user and starting the reply.
The Starlink Test — Real-World Validation
SpaceX tested Grok Voice Think Fast 2.0 on Starlink customer support and sales calls before the public launch, reporting improved sales conversion rates and support containment. This is more informative than synthetic benchmark scores for enterprise buyers: Starlink has millions of subscribers, generates high call volume, and operates across varied network conditions and environments. The model's reported 10× improvement in noisy setting transcription is directly relevant to phone support use cases where callers are not in a quiet office. That said, SpaceX is a related entity to xAI — this is not an independent third-party test. The results are consistent with the technical claims but should be treated as operator-reported, not independently verified.
The $0.08/Min Pricing — What Changed and What It Means
Think Fast 1.0 launched at $0.05/min. Think Fast 2.0 launches at $0.08/min — a 60% price increase per minute of audio. As Basenor's analysis notes, the cost jump from $0.05 to $0.08/min is worth modeling against your usage volume before the August 5 automatic migration happens. For a voice agent handling 10,000 minutes per month: Think Fast 1.0 = $500/month; Think Fast 2.0 = $800/month. For a team running 100,000 minutes/month: $5,000 vs $8,000/month. Whether the transcription accuracy, token efficiency, and agentic improvements justify that 60% cost increase depends on your specific use case and current error rates.
| Competitor | Pricing model | Approx. cost |
|---|
| Grok Voice TF 2.0 | Per minute of audio | $0.08/min |
| Grok Voice TF 1.0 | Per minute of audio | $0.05/min |
| OpenAI GPT-Realtime-2 | Per token (audio) | $32/M input + $64/M output audio tokens |
August 5 Migration — What Developers Need to Do
August 5 deadline: On August 5, 2026, grok-voice-latest auto-migrates from Think Fast 1.0 to Think Fast 2.0. No action needed to upgrade. To stay on v1.0, pin grok-voice-think-fast-1.0 in your API calls before August 5.
Prompt compatibility: According to SpaceXAI, 2.0 works with existing prompts without changes — no prompt engineering required to migrate.
Cost modelling before August 5: Calculate your monthly call volume in minutes × $0.08 and compare to your current × $0.05 bill. If the 60% cost increase is material, run 2.0 on a subset of calls first to validate that the accuracy and efficiency gains justify the premium.
Access: Available now via the xAI API at console.x.ai and through Voice Agent Builder (no-code). Add ~$0.01/min for telephony if using a provisioned phone number.
Who Should Use Think Fast 2.0
High-volume customer support voice agents: The 10× noisy environment improvement and 1.4× transcription gain are directly relevant to call centre replacement use cases. If your current v1.0 deployment has transcription errors on phone calls from noisy environments, 2.0 addresses the primary failure mode at 60% more cost per minute.
Voice agents with complex tool workflows: Tool calls that begin before the first sentence ends mean agentic voice agents feel significantly more responsive. For applications where the agent needs to look up customer data, check inventory, or call an API during a conversation — the tool call latency improvement is meaningful.
Stay on v1.0 if: Your current transcription accuracy is acceptable and the 60% cost increase from $0.05 to $0.08/min is not justified by the gains for your specific call types. Pin grok-voice-think-fast-1.0 before August 5 to prevent auto-migration.
Sources: SpaceXAI official announcement · Basenor developer guide · CryptoBriefing enterprise analysis · DogeDesigner specs summary on X · Related: Grok Aurora voice mode consumer guide → · Grok 4.5 full review → · Claude Voice vs ChatGPT Voice →