Elon Musk: Grok Bot Can Now Speak
Elon Musk, head of xAI, retweeted the product account stating that the grok bot can now talk. The robot opened early testing to the public on August 11, positioned as an AI teammate that can log into user tools and complete tasks, previously focused on text interaction.
The voice layer connects to grok voice, with the current generation being the think fast 2.0 released on July 29. This integration compresses listening, thinking, and speaking into the same model, rather than stitching together speech recognition, language models, and synthesis. The official latency is about 0.70 seconds, utilizing a bidirectional websocket for simultaneous audio input and output. The interface supports over 25 languages, accents, and interruptions, along with voice cloning.
The grok bot testing initially launched on desktop and iOS, with Android following in early September, targeting supergrok heavy and some high-tier cursor subscriptions. The product form builds robots based on roles: research, writing, consulting, inbox management, and sales, continuing to operate after the computer is shut down, then returning the finished product. Early internal testing expanded from company prototypes, with Musk stating that foundational issues needed to be addressed before expanding access after releasing grok 4.6.
The same voice system has been implemented in other areas. Grok voice is available in Tesla vehicles; Starlink customer service and sales handle over 15,000 calls daily. For developers, grok voice is available on fal.ai, charging approximately $0.00083 per second of audio, with continuous one-hour conversations costing about $3; the xai console also offers a voice agent builder and clones of about one-minute samples. The flagship voice has expanded from an initial 5 to an additional 21, covering customer service, character roles, explanations, and more.
This is not the first time voice has appeared in the grok application. The consumer-side voice mode has long been integrated into the grok app and x premium subscriptions, with evaluations showing this voice stack performing well on tau voice bench and big bench audio voice reasoning. The bot's ability to speak effectively combines "a capable teammate" and "a voice that can make calls" into the same product, rather than opening another chat window.
The market mechanism transforms voice from a demonstration feature into an execution entry point. Buyers are individuals and customer service scenarios willing to hand over tool accounts to agents and issue commands verbally; sellers are xai, amortizing training costs across the same grok voice used in bots, vehicle systems, Starlink calls, and developer APIs. Beneficiaries include the low-latency voice stack and existing subscription layers; those under pressure are pure chatbots that still rely on typed prompts and cannot log into external tools. The funding flow is based on call minutes and agent hours, rather than per-message billing in chat boxes.
Source: Public Information
ABAB AI Insight
Musk has separated grok from the chatbots on x into three lines: voice companionship within applications, hands-free assistants in vehicles, and bots that can log into tools. The voice mode will first be available to high-tier subscriptions by the end of 2025, with think fast 1.0 and 2.0 launched in April and July 2026, respectively, before connecting the same engine to Starlink traffic. The bot only began public testing in August and started speaking in September, prioritizing enabling agents to use the software before allowing users to command agents verbally. This contrasts with OpenAI's approach of integrating voice into ChatGPT and then creating agents: xai first develops capable teammates, then adds ears and mouths.
The capital path involves monetizing a single voice model across multiple applications. Vehicle systems and Starlink provide real noise, accent, and interruption data; fal.ai and APIs sell by the second to developers; the bot connects voice to already authorized email, browsers, and CRMs. Funding is not for "better sound" alone but to allow the same reasoning stack to serve both hardware entry points and agent hours. Cloning voices reduces the cost of switching lines for brand customer service while also introducing identity impersonation risks into the product.
Key comparisons must be made. Amazon has turned Alexa into a voice assistant that anyone can call upon but has yet to deliver an agent capable of logging into enterprise software; OpenAI's real-time and Google's native audio facilitate low-latency conversations, yet tool permissions remain fragmented. Xai treats Tesla and Starlink as live testing grounds, practicing scripts with tens of thousands of calls daily, then embedding the same capabilities into personal bots. The industry phase shifts from "voice as a feature" to "voice as an operating system entry point": whoever can modify work orders and respond to emails while speaking will dominate scenarios where hands are already occupied.
Structural judgment belongs to technological substitution. The first layer of response being replaced is typing prompts and human agents, not the judgment itself. The mechanism is that once end-to-end voice reduces latency to under one second, people are willing to verbalize tasks instead of typing them; once agents can log into tools, voice upgrades from an interface to a permission channel.
ABAB News · Cognitive Laws
First, be capable of work; speaking becomes meaningful afterward.
The same voice system amortizes costs in vehicles and customer service.
Voice becomes an entry point; permissions are the product.