SpaceXSI released Grok Voice 3.0 on October 14, a speech-to-speech model aimed at businesses that run voice agents for customer support and sales calls. SpaceXSI is the AI business of SpaceX, formerly known as xAI. On the Artificial Analysis Speech-to-Speech Index, the model scored 86.9 and ranked first, ahead of OpenAI's GPT-Live-1 at 83.2 and Google's Gemini 3.8 Live at 82.6.
The name has changed several times in a short period. SpaceX acquired xAI in February, the unit was renamed SpaceXAI in July, and it took the SpaceXSI name earlier in October, according to SiliconANGLE. The product is what developers will judge, and the benchmark claims are the place to start.
Grok Voice 3.0 benchmark results.
Beyond the headline index score, Grok Voice 3.0 reached a 94.8% task success rate and an Elo rating of 1070 in the Speech Agent Arena, the highest of the four models MarkTechPost compared. It scored 98.5% on Big Bench Audio, which ranks it fourth on that test behind StepAudio 3 Realtime. The model improved on Grok Voice 2.0, which scored 81.3 on the same index, but it did not win everywhere. OpenAI led on the tau-Voice test for customer support calls with accents and background noise, 74.5% to 71.6%.
Speed claims.
xAI says the model responds in under one second and describes that as class-leading. It also says its time to first audio is about 40% lower than GPT-Live-1 at its medium setting. MarkTechPost's figures show 0.81 seconds for Grok Voice 3.0 against 1.34 seconds for GPT-Live-1. The same comparison shows Grok Voice 2.0 was faster, at 0.70 seconds, so the gain is relative to rival models rather than to the earlier version.
What it costs developers.
Pricing stays at $0.08 per minute, or $4.80 per hour of audio, the same rate as Grok Voice 2.0. Developers who use the grok-voice-latest alias were switched to the new model automatically and do not need to change their code. On Big Bench Audio, the cost per task came to about 2.5 cents, compared with 3 cents for GPT-Live-1 and 5.5 cents for GPT-Realtime-2.1. MarkTechPost cautions that list prices do not compare neatly, because OpenAI charges separately for backend usage and Google bills audio and text tokens differently.
Company claims and independent results.
Several figures come from SpaceXSI rather than an outside tester. The company says Tesla vehicles handle millions of Grok Voice conversations each day. It also says its Starlink voice agent answers roughly 20,000 customer calls a day, with about 70% resolved without a human, and that one in four sales calls ends in a purchase. Most of the benchmark numbers come from Artificial Analysis or from ServiceNow's open-source EVA-Bench, which covers more than 200 enterprise scenarios and gave Grok the highest accuracy score, at 74%.
What is still unknown.
The reporting does not describe the testing conditions in detail, and the call-volume and resolution figures have not been independently checked. Developers building voice agents for noisy or accented callers will need their own trials, since the tau-Voice result suggests Grok is not the leader there. A practical first step is to run a sample of real call recordings through both models and compare response delays. Teams that schedule those trials across several offices can use the free time zone converter at /en/time-zone-converter to find shared working hours.