SpaceXAI ships Grok Voice Think Fast 2.0 at a 60% price premium

SpaceXAI released Grok Voice Think Fast 2.0 on July 29th, a speech-to-speech model priced at $0.08 per minute of audio, a 60% increase from the $0.05 rate for Think Fast 1.0. The model, available via API and Voice Agent Builder, claims faster response times and improved accuracy, with an 82.9% score on the Artificial Analysis Speech-to-Speech Index. Existing customers face an August 5th deadline to pin the older model before the default alias switches to the higher-priced version.

Elon Musk @elonmusk https://x.com/elonmusk 's SpaceXAI released Grok Voice Think Fast 2.0 on July 29th, pitching a faster and more accurate speech-to-speech model for developers building customer support, sales and other real-time voice agents. SpaceXAI made the release available through the API and Voice Agent Builder https://x.ai/voice , according to an X thread https://x.com/spacexai/status/2082529280341553209?s=46 and a more detailed product announcement https://x.ai/news/grok-voice-think-fast-2 . The model costs $0.08 per minute of audio, up 60% from the $0.05 rate for Think Fast 1.0. Text input adds $0.004 per request. Musk began xAI's operations in March 2023 and led it until SpaceX acquired xAI on February 2nd, 2026 https://x.ai/news/xai-joins-spacex . A June SpaceX prospectus https://content.spacex.com/cms-assets/FINAL Documents%20and%20Updates/SpaceX%20-%20EU%20Prospectus%20%28Approved%20by%20Bafin%29%20-%20June%205%2C%202026.pdf lists Musk as SpaceX's CEO, chief technical officer and chairman. It also traces his path through Zip2 and PayPal and notes his physics and business degrees from the University of Pennsylvania. Grok now sits inside a SpaceX segment that combines AI models, consumer and enterprise software, the X platform and computing infrastructure. The timing shows SpaceXAI moving Grok Voice toward commercial call handling. SpaceXAI launched the Voice Agent Builder https://x.ai/news/grok-voice-agent-builder in beta on July 1st, giving developers and operators a visual interface for configuring agents with telephone numbers, knowledge retrieval, tools, guardrails, recordings and transcripts. Customers can connect existing numbers over SIP or use WebSockets to bring the agents into their own applications. The upgrade and the migration deadline SpaceXAI says Think Fast 2.0 reasons while producing speech, allowing it to begin answering as it works through a request. SpaceXAI also says the model uses 40% as many reasoning tokens as Think Fast 1.0 at the median, which should let tool calls execute earlier during a response. The release carries a migration decision for existing customers. On August 5th, the grok-voice-latest alias will move automatically from Think Fast 1.0 to Think Fast 2.0. Developers who want to retain the cheaper model must pin grok-voice-think-fast-1.0 before that switch. That default migration gives SpaceXAI a direct route to move existing usage onto the higher-priced model. At sustained volume, the difference is material: one hour of audio rises from $3 to $4.80 before text-input charges. SpaceXAI is asking customers to accept that increase in exchange for better task completion, transcription and conversational timing. SpaceXAI claims Think Fast 2.0 achieved an 82.9% score on the Artificial Analysis Speech-to-Speech Index, compared with 75.7% for Think Fast 1.0 and 79.1% for OpenAI's GPT-Realtime-2.1 High. Its claimed time to first audio fell from 1.25 seconds to 0.70 seconds, while its score on the customer-service-focused tau-voice benchmark rose from 52.1% to 56.5%. The Artificial Analysis leaderboard https://artificialanalysis.ai/speech-to-speech corroborates those figures, while putting them in a broader field. Alibaba Cloud's Qwen Audio 3.0 Realtime Plus leads the overall index at 84.1%, ahead of Grok's 82.9%. Deepslate Opal and Google's Gemini 2.5 Flash Native Audio Dialog also produce initial audio faster, according to the same leaderboard. Grok's strongest result is agentic performance: its 56.5% tau-voice score leads the models listed by Artificial Analysis. SpaceXAI separately claims that an internal evaluation of thousands of short phrases across 24 languages showed Think Fast 2.0 cutting transcription errors by 1.5x to 2x relative to Deepgram Nova 3 and ElevenLabs Scribe v2. It says the advantage widened to roughly 10x under background noise and telephone compression. Those transcription figures come from SpaceXAI's own evaluation rather than the independent leaderboard. A full-stack voice product SpaceXAI is selling an integrated speech-to-speech path instead of assembling separate speech recognition, language and speech-generation services. Its Voice Agent Builder wraps that model with the infrastructure needed to answer telephone calls, search internal documents, invoke external tools, transfer callers and review completed conversations. That puts SpaceXAI into direct competition with OpenAI's realtime models and ElevenLabs Agents https://elevenlabs.io/docs/eleven-agents/overview , as well as voice infrastructure from Google and Deepgram. The buying criteria have moved beyond whether an agent can hold a convincing demo conversation. Developers deploying agents into call centers have to measure latency, interruption handling, transcription under poor audio, successful tool execution and the cost of every minute. SpaceXAI has a ready-made proving ground inside Musk's other operations. SpaceXAI says it tested Think Fast 2.0 on Starlink's customer telephone line and recorded higher sales conversion and support containment. The release turns that internal deployment into a developer product, pairing SpaceX's distribution and computing resources with a model priced for production call volume.