SpaceXSI launched Grok Voice 3.0 on October 14, 2026, rolling out the AI customer support model with a reported top score of 86.9 on Artificial Analysis’ Speech-to-Speech Index, techrepublic.com reported. The release sets up an immediate test for IT teams evaluating automated voice agents against existing enterprise stacks.
Benchmark Scores and Latency Trade-Offs in Enterprise Testing
The updated speech model outpaces several established competitors on raw metric scores, though performance varies significantly across specific audio tests. According to techrepublic.com, Artificial Analysis scored Grok Voice 3.0 at 86.9, placing it slightly ahead of OpenAI’s GPT-Live-1 (Astra medium configuration) at 83.2 and Google’s Gemini 3.8 Live with Extended Thinking set to High at 82.6. The predecessor, Grok Voice 2.0, previously scored 81.3.
Introducing Grok Voice 3.0, the world’s most intelligent speech-to-speech model.It excels at resolving customer calls, responds in under a second, and costs less than other frontier models. pic.twitter.com/cfDRrkQPEc
Additional evaluations show the model achieving 97% on Full Duplex Bench, which measures natural handling of pauses, interruptions, and turn-taking. Its speech reasoning score reached 98.5% on Big Bench Audio. However, Grok Voice 3.0 did not capture every category. When evaluated on τ-Voice—a benchmark assessing AI agents on customer support calls that feature accents, background noise, and interruptions—the model recorded 71.6%, coming in behind the 74.5% achieved by GPT-Live-1 in its Astra medium configuration.
Response latency introduces another operational variable. Artificial Analysis measured Grok Voice 3.0’s average time to first audio at 0.81 seconds. While faster than GPT-Live-1’s Astra medium configuration at 1.34 seconds, it marks a regression from Grok Voice 2.0, which recorded 0.70 seconds.
Deployment Scale and Operational Integration at SpaceXSI and Tesla
SpaceXSI deploys the technology internally, reporting that Starlink uses Grok Voice to handle approximately 20,000 customer calls daily across multiple languages, resolving roughly 70% without human intervention, techrepublic.com reported. The system executes functions including customer identity verification, hardware troubleshooting, account credit issuance, and replacement equipment scheduling. The company also states that one in four sales calls handled by the system ends in a purchase before conversation completion.
Beyond call centers, the model powers millions of daily conversations in Tesla vehicles for hands-free requests such as checking calendars and managing email. It also drives Grok Bot’s voice mode. Furthermore, techrepublic.com noted a 74% accuracy score on ServiceNow’s EVA-Bench, which evaluates voice agents across more than 200 enterprise scenarios.
Nitin Aggarwal, ServiceNow’s senior director of product management for Agentic Evals and RAI, stated in SpaceXSI’s announcement that the model pairs top accuracy with a natural, responsive conversation.
API Pricing Structure and Development Implications
As reported by techrepublic.com, Grok Voice 3.0 remains at the same pricing tier as Grok Voice 2.0, accessible through SpaceXSI’s Voice Agent API for a base audio rate of $4.80 per hour, or $0.08 per minute. Developer documentation also specifies a $0.004 charge per billable text input message. Default voice-activity-detection sessions are billed for the full session duration, whereas push-to-talk sessions incur charges only for audio sent and received.
Developers utilizing the grok-voice-latest alias experienced an automatic transition to the new version. Technical teams using this endpoint must rerun representative support calls to verify tool use, escalation behavior, and response timing. Organizations evaluating total deployment costs must factor in integration, monitoring, and human escalation expenses, as a lower per-minute model charge offers limited savings if callers require repeated attempts or staff must manually correct agent actions.
