[DigitalToday reporter Jinju Hong] SpaceXSI has launched its Grok Voice 3.0 voice AI model to target the enterprise voice agent market. It outperformed models from OpenAI and Google in an independent voice-to-voice evaluation, and the company also highlighted competitiveness in processing costs and response speed.

SiliconANGLE reported on Oct. 14 that Grok Voice 3.0 ranked first with a score of 86.9 on the voice-to-voice index from AI model evaluation firm Artificial Analysis. OpenAI's GPT-Live-1 scored 83.2, followed by Google's Gemini 3.8 Live at 82.6.

The evaluation was based on reasoning tests, task success rates and blind call-listening assessments. SpaceXSI plans to deploy Grok Voice 3.0 in the voice agent market for handling enterprise tasks such as customer support and sales calls.

It also showed strength in handling natural conversation. Grok Voice 3.0 scored 97 percent on the full-duplex bench, which measures how naturally it responds to silence and interruptions during a conversation.

It did not lead in every assessment. In the Tau-Voice test, which assumes customer support calls with strong accents and background noise, OpenAI's model narrowly led with 74.5 percent versus Grok Voice 3.0's 71.6 percent. SpaceXSI said the time to its model's first voice response was about 40 percent shorter in that test.

It also stressed cost competitiveness. Under Artificial Analysis' Big Bench Audio standard, the cost to process 1 task was about 2.5 cents for Grok Voice 3.0. GPT-Live-1 was 3 cents, and OpenAI's GPT-Realtime-2.1 was 5.5 cents. The company said enterprises could reduce processing costs when adopting voice AI for large-scale customer response operations.

The model also showed performance in evaluations targeting enterprise work scenarios. In ServiceNow's open-source evaluation tool EVA-Bench, it posted 74 percent accuracy across more than 200 enterprise scenarios. It also scored higher than competing models in an evaluation that considered both accuracy and customer experience.

Nitin Agrawal (니틴 아가왈), a senior product management director at ServiceNow, assessed that Grok Voice 3.0 showed both high accuracy and natural, responsive conversation capabilities.

Service applications are also under way. Tesla vehicles are conducting millions of Grok Voice conversations a day. Drivers can use voice to check schedules and reply to emails, and tasks such as restaurant reservations can be handled through the Grok bot. The new model has also been applied to the Grok bot's voice mode.

Starlink is handling about 20,000 customer calls a day in multiple languages using Grok Voice. SpaceXSI said about 70 percent of these are resolved without human intervention. It also said 1 in 4 sales calls leads to a purchase before the call ends.

The launch is a follow-on product after SpaceX's AI organisation restructuring. SpaceX acquired xAI in February and changed the AI organisation's name to SpaceXAI in July. Earlier this month it changed it again to its current name, SpaceXSI.

Grok Voice 3.0 is currently available through the voice agent API. Pricing is 8 cents per minute, or $4.80 per hour, the same as the previous version, Grok Voice 2.0. Developers who previously used the 'grok-voice-latest' model alias will be automatically switched to the new version without additional modifications.