On Wednesday, February 14th, US Eastern Time, SpaceX AI unit SpaceXAI released its next-generation voice model Grok Voice 3.0, emphasizing low latency, high task success rates, and enterprise-grade voice agent applications.

SpaceX touted Grok Voice 3.0 as the "world's most intelligent speech-to-speech model," saying it excels at handling customer calls, can respond to customers in under one second, costs less than frontier industry models, and can already autonomously resolve 70% of customer service calls for SpaceX's Starlink business. Developers can access the model starting today through an API and voice agent building tools, priced at $0.08 per minute.

Independent AI evaluation platform Artificial Analysis further showed that Grok Voice 3.0 ranked first on its speech-to-speech model composite index with a score of 86.9, beating multiple competitors including OpenAI's GPT-Live-1 (Astra, medium) and Google's Gemini 3.8 Live Extended Thinking. However, the model is not a comprehensive leader in user preference and certain agent task tests, and its response speed has declined compared with the previous generation.

This launch highlights a new direction in AI competition: model makers are no longer only competing over text generation and reasoning capabilities, but are also vying for the voice agent market where systems can directly answer calls, understand voice commands, invoke business tools, and complete real tasks. For enterprise customers, response speed, task completion rates, and per-minute calling costs will become key metrics for measuring the commercial value of such products.

Top composite score, 94.8% task success rate

According to test results published by Artificial Analysis, Grok Voice 3.0 scored 86.9 on the Speech to Speech Index, ranking first on its debut. The index comprehensively examines voice reasoning, agent execution capability, user preference, and task success rate.

For comparison, GPT-Live-1 (Astra, medium reasoning configuration) scored 83.2, Google Gemini 3.8 Live Extended Thinking (high reasoning configuration) scored 82.6, and GPT-Live-1 (Sol, low reasoning configuration) scored 80.1. Grok Voice 3.0 improved by 5.6 points over the previous generation Grok Voice 2.0's 81.3 and by 13.2 points over Grok Voice 1.0's 73.7.

On specific metrics, Grok Voice 3.0's performance was mixed:

Leading task success rate. In the Speech Agent Arena test, the model's task success rate reached 94.8%, ranking first and slightly above Grok Voice 2.0's 94.6%.

Third in user preference. Its Elo score was 1070, below Gemini 3.1 Flash Live Minimal's 1100 and Gemini 3.8 Live's 1088, but above GPT-Live-1 (Sol, low configuration)'s 1057.

Second in voice agent execution capability. In the Tau Voice test, Grok Voice 3.0 scored 71.6%, behind only GPT-Live-1 (Astra, medium configuration)'s 74.5% and above Gemini 3.8 Live Extended Thinking (high configuration)'s 68.6%. By comparison, Grok Voice 2.0 scored 56.5%.

Outstanding audio reasoning performance. In the Big Bench Audio test, Grok Voice 3.0 scored 98.5%, exceeding Google's related model's 97.7% and the previous generation Grok Voice 2.0's 97.2%.

The above results indicate that Grok Voice 3.0's advantages are especially evident in task execution, audio reasoning, and overall performance, but it still lags behind Google's two models in subjective user preference.

It should be noted that these results come from Artificial Analysis's specific testing framework and cannot be directly equated with performance in all real customer service scenarios. Actual results still depend on business processes, external tools, phone systems, and the complexity of customer issues.

First response about 0.81 seconds, slightly slower than previous generation

For voice agents that require real-time interaction, response latency is an important metric affecting user experience.

Artificial Analysis's tests showed that in the Big Bench Audio test, Grok Voice 3.0's average time to first audio response was 0.81 seconds, faster than GPT-Live-1 (Sol, low configuration)'s 1.24 seconds, GPT-Live-1 (Astra, medium configuration)'s 1.34 seconds, and Google Gemini 3.8 Live Extended Thinking (high configuration)'s 1.35 seconds.

But compared with the previous generation Grok Voice 2.0's 0.70 seconds, the new model's time to first audio response actually increased by 0.11 seconds. This means that although Grok Voice 3.0 improved in overall capability and multiple task metrics, it did not surpass its predecessor on all speed indicators.

SpaceX said in its official introduction that Grok Voice 3.0 achieved frontier-level performance on voice agent tasks while reducing latency by 40%. This statement does not fully correspond to the specific first-response times in third-party tests; it may involve different latency definitions or test scenarios, and the two should not be directly treated as the same metric.

For enterprise deployment, what truly matters is not only how quickly the model starts speaking, but also whether it can continuously understand users across multi-turn conversations, correctly invoke tools, and ultimately complete business tasks such as refunds, account inquiries, and fault diagnosis.

Charging $0.08 per minute, targeting the enterprise voice agent market

Grok Voice 3.0 is already available through SpaceXAI's real-time API and can be invoked in Voice Agent Builder. Official pricing is $0.08 per minute, or $4.80 per hour, unchanged from the previous generation Grok Voice 2.0.

At this price, if a single call lasts 5 minutes, the audio invocation cost would be about $0.40; if 100,000 minutes of audio are processed cumulatively per month, the cost would be about $8,000. The above estimates are based only on the published audio rate; actual project budgets also need to consider possible costs from other text input, external tools, and phone lines.

SpaceXAI also sets grok-voice-latest to point to Grok Voice 3.0 by default. This means existing developers using that dynamic model alias will automatically switch to the new version without modifying the model name; the original Grok Voice 2.0 has been marked as deprecated. The company said the new model uses the existing API, so developers do not need to re-adapt the entire interface just to upgrade.

On price, Grok Voice 3.0 is not the cheapest option among all competitors. Artificial Analysis estimates its hourly input audio cost at $4.80, the same as the previous generation; GPT-Live-1 (Sol, low configuration) is $4.47, Google Gemini 3.8 Live Extended Thinking (high configuration) is $3.50, and GPT-Live-1 (Astra, medium configuration) is $5.83. By comparison, GPT-Realtime-2.1 high configuration's corresponding cost is $10.75.

Therefore, Grok Voice 3.0's selling point is not simply a low price, but rather improving its composite evaluation score and task completion capability while maintaining the previous generation's price, and launching it as a voice agent model developers can directly invoke.

Starlink customer service already autonomously handles 70% of calls

SpaceX provided an internal business application case for the new model: officially, Grok Voice 3.0 can now autonomously resolve 70% of Starlink customer service calls.

This figure is an important signal for measuring the commercial value of voice agents. If the model can handle common account inquiries, connection troubleshooting, and service issues without human intervention, enterprises can reduce manual customer service workload and concentrate staff on complex complaints and cases requiring human judgment.

However, SpaceX did not disclose in the above introduction the statistical period, call sample size, distribution of issue types, or the specific criteria for determining "autonomous resolution" for this 70% figure. Therefore, this number should be regarded as an internal application result published by the company, not an independently third-party verified industry average, and it cannot be used to directly calculate the specific costs enterprises can save.

From a product positioning perspective, Grok Voice 3.0 targets voice agents capable of executing tasks, not just traditional speech synthesis systems that convert text to speech. Such models need to understand user intent in real-time conversation, maintain context, handle ambiguous requests, and invoke external tools when necessary to complete operations. Their actual value ultimately depends on whether they can reliably solve problems, not merely whether their answers sound natural.

For SpaceX, Starlink customer service also provides a real business environment for application scenarios: the company can both use its own business to validate the model and open related capabilities to other enterprises and developers through the API.

Voice AI competition shifts toward actual delivery capability

The release of Grok Voice 3.0 reflects that the focus of competition in the voice AI market is shifting from a single emphasis on voice naturalness and response speed to a balance among comprehensive reasoning, task execution, latency, and cost.

Artificial Analysis's results show that Grok Voice 3.0 leads in the composite index and task success rate, but ranks third in the user preference list, and its time to first audio response is also slower than the previous generation. This shows that a single composite ranking cannot fully reflect the relative advantages and disadvantages of different products in actual applications.

For enterprise customers, choosing a voice model also requires examining stability in long calls, completion rates for complex tasks, error-handling mechanisms, and overall cost under real business traffic. Especially in industries with high requirements for accuracy and compliance, such as finance, telecommunications, and healthcare, higher automation rates do not mean that manual review and escalation mechanisms can be completely eliminated.

As SpaceXAI opens Grok Voice 3.0 to developers, competition will further extend into enterprise workflows and customer service systems. Whoever can continuously complete real tasks at controllable cost, rather than merely scoring high on benchmarks, will be more likely to turn voice agents into scaled commercial products.