Grok Voice 3.0 tops Artificial Analysis voice benchmark

SpaceXSI and its affiliated site SpaceXAI have introduced Grok Voice 3.0, a new speech-to-speech model positioned for real-time conversations and enterprise customer service. The main headline around the launch is a third-party ranking: according to results cited from Artificial Analysis, Grok Voice 3.0 scored 86.9 on the group’s Speech-to-Speech Index and placed ahead of tested models from OpenAI and Google. That makes the release notable, but the published material also shows why the result should be handled carefully. The benchmark aggregates several dimensions, and some individual tests paint a more mixed picture.

What Grok Voice 3.0 is meant to do

Both sources describe Grok Voice 3.0 as a speech-to-speech model, meaning it takes spoken input and produces spoken output directly. The model is presented as supporting live dialogue, interruptions, turn-taking and voice-agent tasks. SpaceXAI says it has class-leading latency at under one second, while iThome reports that SpaceXSI claims response times can fall below one second. In both cases, that latency claim comes from the company side, not from a full independent test summary.

The launch is framed around customer support and voice agents rather than only consumer chat. SpaceXAI says Grok Voice 3.0 is already used across Tesla and SpaceX operations, including Tesla vehicles and Starlink call centers. Those usage figures and operational outcomes are company claims reported in the sources. They are relevant because they suggest the model is being positioned as an applied product, not just a research demo, but they are not the same thing as independently verified market performance.

What Artificial Analysis published

The most concrete external measurements in the sources come from Artificial Analysis. According to the material cited by iThome and SpaceXAI, Grok Voice 3.0 leads the Speech-to-Speech Index with a score of 86.9. iThome says this is higher than Grok Voice 2.0 at 81.3, OpenAI GPT-Live-1 at 83.2, and Google Gemini 3.8 Live at 82.6 in the October 13 results it references.

The index is described as combining multiple factors. The sources say it draws on speech reasoning, human preference, task success in blind head-to-head calls, and a measure referred to as τ-Voice. SpaceXAI also says Grok Voice 3.0 scored 97% on Full Duplex Bench, which it describes as a measure of how naturally the model handles pauses, turn-taking and interruptions. iThome similarly reports a 97% result on tests around conversation pauses, turn-taking and user interruptions.

• Speech-to-Speech Index score: 86.9
• Speech reasoning accuracy: 98.5%
• Task completion rate: 94.8%
• Full Duplex Bench or equivalent conversational handling score: 97%
• τ-Voice task success in noisy, accented, interruption-heavy support scenarios: 71.6%

Cost is another part of the published comparison. iThome says Artificial Analysis measured an average cost of $0.025 per speech reasoning task for Grok Voice 3.0, below $0.030 for GPT-Live-1 and $0.055 for GPT-Realtime-2.1. SpaceXAI also says the model is cheaper to run than comparable models in Artificial Analysis’s Big Bench Audio run. These figures matter because enterprise buyers usually weigh quality, speed and cost together rather than in isolation.

Where the benchmark is strong, and where it is limited

The benchmark matters because it is not presented as an internal company test. Both sources describe Artificial Analysis as an independent or third-party evaluator, and the published dimensions go beyond a single quality score. That gives the launch more substance than a vague marketing claim. The use of blind head-to-head calls, task success, speech reasoning and conversation flow suggests the benchmark is trying to capture several aspects of voice-agent performance.

Still, the same published numbers also show why a top ranking should not be treated as absolute proof that one model is best in every deployment. A combined index inevitably reflects methodological choices: what tasks are included, how they are weighted, and which trade-offs matter most. The sources do not provide the full benchmark design, so readers can see the outcomes but not every detail behind them.

More importantly, one of the subtests points to a limitation. In τ-Voice, described as simulating support calls with diverse accents, background noise and interruptions, Grok Voice 3.0 scored 71.6%. iThome says that is slightly below OpenAI GPT-Live-1 at 74.5%, even though Grok Voice 3.0 had 40% lower latency. SpaceXAI characterizes the result as comparable with GPT-Live-1 and emphasizes the latency advantage. Those two descriptions are not necessarily contradictory, but they frame the same comparison differently. The key point is that Grok Voice 3.0 does not lead every published measure in the sources.

That nuance is important for enterprise voice systems. A model can rank first overall while trailing a rival in a scenario that looks closer to a difficult real call-center environment. For some buyers, lower latency may outweigh a small gap in success rate. For others, especially where accents, noise and interruptions dominate, the trade-off may be judged differently. The benchmark helps identify that trade-off, but it does not settle it for every use case.

EVA-Bench and the enterprise angle

Beyond Artificial Analysis, the sources also cite EVA-Bench, which they describe as a ServiceNow benchmark for enterprise voice agents spanning more than 200 customer-service scenarios. According to both sources, Grok Voice 3.0 achieved the top accuracy score there at 74%. SpaceXAI adds that the model is the best-performing one when jointly scored across accuracy and customer experience, and includes a favorable quote about a new Pareto frontier on that benchmark.

That quote is positive, but it comes from the launch material reproduced by SpaceXAI, so it should be read as part of the company’s presentation unless independently corroborated. Even so, the repeated reference to EVA-Bench in both sources strengthens the idea that Grok Voice 3.0 is being judged not only on conversational fluency but also on support-style task handling. This is central to the product story: the model is being sold as infrastructure for automated service interactions.

Availability, pricing and what companies should watch

SpaceXAI says Grok Voice 3.0 is available now in the Voice Agent API at $0.08 per minute, the same price as Grok Voice 2.0, and that applications using the grok-voice-latest model name were switched over automatically. iThome reports the same pricing and automatic upgrade path. The launch materials also say Grok Voice supports millions of daily conversations in Tesla vehicles and around 20,000 calls a day in Starlink customer service, with roughly 70% resolved without human intervention and one in four sales calls ending in a purchase. These operational figures are company claims in the sources, not third-party audit results.

For English-speaking businesses, the practical takeaway is not simply that Grok Voice 3.0 came first in one ranking. It is that a newly released voice model is being presented with published third-party scores on quality, speed and cost, plus API availability at unchanged pricing. That may make it worth evaluating for customer support or voice-agent workflows, especially if low latency is a priority. But companies should compare it against their own call types, accents, noise conditions and success metrics, because the same published results show that leading an overall benchmark does not guarantee the best outcome in every real support scenario.