Grok Voice Think Fast 2.0 at a glance

What to consider

• The model can mishear names, numbers, addresses, dates, account details, accents, quiet speech, overlapping speakers, or low-quality telephony audio.
• It can hallucinate policies, diagnoses, legal conclusions, prices, availability, tool results, confirmations, or whether an action succeeded.
• Headline benchmark results in the launch post come from a particular benchmark provider and setup; they do not establish performance for a specific production call flow.
• The Starlink sales-conversion and support-containment claim is vendor-reported A/B-test evidence without enough public detail to predict another organization's outcome.
• Reasoning while speaking can create a polished response before the underlying tool result is fully known, so clients must prevent premature factual claims or stale follow-ups.
• Tool calls turn an audio mistake or prompt-injection attack into a possible real-world action.
• Caller speech, background audio, retrieved documents, web pages, tool output, and transferred context can all contain malicious or conflicting instructions.
• Voice and caller ID are not reliable identity factors and can be spoofed or cloned.
• Recording, transcription, outbound calling, automated-decision, biometric, accessibility, and AI-disclosure rules vary by jurisdiction and use case.
• Custom voices create consent, impersonation, ownership, and fraud risks and are currently restricted geographically.
• The $0.08 rate applies to audio sent or received, so a two-sided conversation can cost more than wall-clock call duration suggests.
• Carrier, SIP, phone-number, tool-provider, retrieval, application, observability, and human-escalation costs are not included in the model rate.
• The direct model page lists 10 concurrent sessions per team and a 120-minute maximum, which may require a limit increase or capacity plan.
• The current model page lists only us-east-1 for direct availability; geography, residency, and latency need confirmation for the target audience.
• Session resumption is opt-in and documented history expires after 30 minutes of inactivity, so applications must own durable business state separately.
• SpaceXAI says realtime audio is not stored, while its general API security FAQ describes 30-day default retention for API requests and responses; teams should confirm how transcripts, text events, tool payloads, and logs are treated for their configuration.
• Zero Data Retention changes or disables stored-data features and must be validated at the team and application level.
• The `grok-voice-latest` alias can change behavior without a code edit; production systems need a pinned fallback and regression tests.
• A voice agent must provide an immediate path to a qualified human for emergencies, vulnerable users, high-impact decisions, authentication failures, and repeated misunderstanding.
• No voice model should autonomously make medical, legal, financial, employment, security, or safety-critical decisions.