xAI Launches Grok Voice 3.0, a New Speech-to-Speech Model for Voice Agents

ByAI Tool Insights Team Hours Categories Generative AI
Grok Voice 3.0 voice AI editorial illustration

xAI announced Grok Voice 3.0 on October 14, 2026, the newest version of its speech-to-speech model for voice agents. The model is available immediately through the company’s Voice Agent API at the same price as its predecessor, $0.08 per minute of conversation, while xAI claims it tops the Artificial Analysis Speech-to-Speech Index with a score of 86.9. The quiet but consequential part of the launch: any application using the grok-voice-latest alias was switched to the new model automatically, with no deploy required on the developer’s side.

Background

The voice-agent API category has become one of the busiest fronts in the AI platform competition. OpenAI’s GPT-Live-1, Google’s Gemini Live line, and xAI’s Grok Voice models are all vying for the developers building phone agents, in-car assistants, and customer-service bots. Grok Voice 2.0, the previous version, scored 81.3 on Artificial Analysis’s Speech-to-Speech Index.

xAI already operates Grok Voice at significant production scale. According to the company, the model handles millions of conversations a day in Tesla vehicles and powers Grok Bot’s voice mode, while Starlink call centers use it to answer 20,000 customer calls every day across multiple languages. xAI says roughly 70 percent of those calls are resolved with no human in the loop, and one in four sales calls ends in a purchase before the caller hangs up. These deployment figures are company-reported and separate from the independent benchmark results.

Key details

Pricing is unchanged. Grok Voice 3.0 costs $0.08 per minute, or $4.80 per hour, plus $0.004 per text input, per xAI’s pricing page. Both voice generations carry those rates, so the upgrade improves measured capability without changing the base price of a fixed volume of conversation.

The API itself is unchanged. Per xAI’s release notes, “grok-voice-latest now points to grok-voice-3.0, so sessions that use the alias are upgraded automatically.” Applications pinned to grok-voice-2.0 must change the model identifier; version 2.0 is deprecated, although the release notes give no retirement date yet. The endpoint, session parameters, and events remain the same, so existing integrations should keep working. The docs show a WebSocket connection to wss://api.x.ai/v1/realtime with the API key in an authorization header, and recommend short-lived ephemeral tokens for client-side applications so the API key is not exposed in the app. Reasoning defaults to high effort, with an option to disable it.

Benchmarks favor the new model, with one caveat. xAI’s announcement cited independent testing by Artificial Analysis dated October 13, 2026, placing Grok Voice 3.0 first on the Speech-to-Speech Index at 86.9, ahead of GPT-Live-1 (Astra, Medium) at 83.2, Gemini 3.8 Live (Extended Thinking, High) at 82.6, and Grok Voice 2.0 at 81.3. The index is an equal-weighted score across four components: speech reasoning from Big Bench Audio, agentic performance from tau-Voice, arena preference, and task success rate.

Across the components, xAI reported 98.5 percent on Big Bench Audio speech reasoning, a Speech Agent Arena preference Elo of 1070, a 94.8 percent task success rate, and 97.0 percent on Full Duplex Bench, which measures how naturally a model handles pauses, turn-taking, and interruptions.

The agentic benchmark tau-Voice, built from simulated customer-support calls, tells a more nuanced story. Grok Voice 3.0 reached 71.6 percent, a 15.1-point jump from Grok Voice 2.0’s 56.5 percent, but GPT-Live-1 (Astra, Medium) still led at 74.5 percent. On latency, xAI claims under one second and 40 percent lower latency than GPT-Live-1 on tau-Voice, while Artificial Analysis independently measured first-audio latency at 0.81 seconds, slightly higher than the 0.70 seconds for Grok Voice 2.0.

xAI also reported that Grok Voice 3.0 achieved the highest accuracy score on EVA-Bench, ServiceNow’s open-source benchmark for enterprise voice agents, at 74 percent, and led when accuracy and customer experience were scored jointly. Nitin Aggarwal, ServiceNow’s Senior Director of Product Management for Agentic Evals and RAI, said in a statement published with the announcement that Grok Voice 3.0 “sets a new Pareto frontier on EVA-Bench.”

Why it matters

For developers, this is a straightforward capability upgrade at a flat price, but the 15.1-point jump in tau-Voice task completion is also a behavioral change worth testing. A model that completes more tasks gets there by different routes: when it yields the turn, when it escalates to a human, when it calls a tool. Prompts tuned against Grok Voice 2.0 may behave differently on 3.0, which is why independent coverage recommends pinning a versioned model ID in production rather than riding the -latest alias.

For the market, the launch confirms that the per-minute price has become a fixed anchor while vendors compete on task completion, latency, and cost per task. GPT-Live-1 still holds the lead on the agentic tau-Voice component, so the benchmark race is far from settled.

Sources

The AI Tool Insights editorial team covers the latest developments in AI tools and the AI industry.

Similar Posts

Leave a Reply

Your email address will not be published. Required fields are marked *