Grok Voice 3.0 Released: More Natural Conversations, Ranked First in Enterprise Customer Service Evaluations, Already in Use at Tesla

xAI has released its next-generation voice model, Grok Voice 3.0, calling it "the world's most intelligent speech-to-speech model." It focuses on real-time calling scenarios such as customer service phone calls: it ranks first on third-party evaluation organization Artificial Analysis's Speech-to-Speech Index, speaks in under 1 second, and keeps the same price as the previous generation. It is already in large-scale use inside Tesla vehicles and Starlink customer service, and is now available to developers via API.

Key highlights:

• Ranked First on the Comprehensive Leaderboard: Scored 86.9 points on the Artificial Analysis Speech-to-Speech Index, higher than GPT-Live-1 (83.2) and Gemini 3.8 Live (82.6).
• Speaks in Under 1 Second: Emits the first sound approximately 0.81 seconds after the caller finishes speaking; in customer service call tests, its accuracy is close to GPT-Live-1 while responding about 40% faster.
• Gets Things Done on Difficult Calls: In simulated customer service calls with accents, background noise, and mid-call interruptions, the resolution rate increased from 56.5% in the previous generation to 71.6%.
• More Natural Conversations: Scored 97% on Full Duplex Bench, which measures handling pauses, turn-taking, and interruptions.
• Ranked First in Enterprise Customer Service Benchmarks: Achieved 74% accuracy on ServiceNow's enterprise voice Agent benchmark EVA-Bench, ranking first.
• Already in Large-Scale Use: Millions of conversations daily inside Tesla vehicles; Starlink customer service handles 20,000 calls daily, with about 70% requiring no human intervention.
• Unchanged Pricing: Voice Agent API costs 0.08 USD per minute, the same price as Grok Voice 2.0; applications using grok-voice-latest have been automatically upgraded.

What Kind of Model Is It

There are two common approaches to voice AI. One is like a three-station pipeline: first converting your speech into text (speech recognition), passing it to a large language model to come up with a text response, and then reading the text aloud (speech synthesis). Each station must wait for the previous one to deliver, causing waiting time to compound; your tone of voice, hesitations, and half-sentence pauses are also lost during transcription.

Grok Voice 3.0 belongs to the other category, a speech-to-speech model: the input is audio and the output is directly audio, without first producing an intermediate text transcript. Models of this kind can start speaking sooner and also support full duplex conversation(like a phone call, where both parties can speak at the same time, and it can hear you and pause when you interrupt). The models compared this time, GPT-Live-1 and Gemini 3.8 Live, also belong to this category, making this a comparison of accuracy, speed, and cost among peer models.

xAI has not disclosed the model architecture or training methods for Grok Voice 3.0; assessing its capabilities relies primarily on benchmark and deployment data.

Where It Improves Over the Previous Generation

Compared to Grok Voice 2.0, changes are concentrated on "difficult calls" and "people being more willing to talk to it":

In routine calls, 2.0 could already handle most tasks, and the task success rate remained virtually unchanged. The improvements in 3.0 lie in challenging situations like accents, noise, and interruptions: taking an extra ~0.1 seconds to speak in exchange for a 15 percentage point gain in resolution rate.

Performance: Accuracy and Speed

τ-Voice tests voice Agents by having them answer simulated customer service calls where callers speak with various accents, background noise is present, and callers interrupt mid-call, evaluating whether the task is ultimately resolved. In the chart below, the vertical axis is resolution rate, and the horizontal axis is "first audio latency" (how long after the caller finishes speaking before the model emits the first sound). The original chart has an inverted horizontal axis—farther right is faster, and closer to the top right is better.

Specific readings (success rates are labeled in the original chart; latencies were not listed numerically and are converted from chart positions, with an error margin of ~0.02 seconds):

• Grok Voice 3.0: 71.6% success rate, speaks in about 0.81 seconds.
• GPT-Live-1 (Astra, Medium): 74.5% success rate, 2.9 percentage points higher than Grok, making it the only model in the chart with higher accuracy; however, it takes about 1.34 seconds to speak, roughly half a second slower. xAI describes Grok's latency as "40% lower," which calculates to about 39.5% lower.
• Gemini 3.8 Live (Ext. Thinking, High): 68.6%, about 1.35 seconds.
• There are 12 other unlabeled gray dots in the chart, all of which are slower or less accurate than Grok Voice 3.0.

In the context of a phone call: when a customer says "my internet is down," one AI responds after 0.8 seconds, while the other takes over 1.3 seconds. That silence from the latter makes it much easier for people to assume the other end didn't hear them. Conversely, GPT-Live-1's extra 2.9 percentage points means roughly 3 more calls resolved out of every 100 such calls. Which one to choose depends on whether a business fears slowness more or errors more.

In Artificial Analysis's comprehensive comparison, Grok Voice 3.0 ranks first across every metric (GPT-Live-1 evaluated at Astra, Medium tier; Gemini 3.8 Live at Ext. Thinking, High tier):

• Comprehensive Index: Artificial Analysis's overall speech-to-speech score, combining speech reasoning, τ-Voice, and preference and resolution rates from blind call tests.
• Big Bench Audio: Spoken reasoning problems, testing "whether it can think correctly after listening."
• Human Preference Elo (Speech Agent Arena): Two models place anonymous calls with humans, who choose which one they prefer, with scores calculated based on wins and losses.
• Task Success Rate: The proportion of tasks ultimately resolved in the same blind call tests.
• Full Duplex Bench: Whether pauses, turn-taking, and interruptions are handled naturally.

The competitors and metrics in this table were chosen by xAI; τ-Voice is presented separately in the scatter plot above, where GPT-Live-1 scored higher.

There is also an external benchmark for enterprise scenarios: ServiceNow's EVA-Bench(ServiceNow's open-source enterprise voice Agent benchmark covering over 200 real enterprise service scenarios, scoring both accuracy and customer experience). Grok Voice 3.0 achieved 74% accuracy, ranking first; combining accuracy and customer experience, it was also the top-performing model. Nitin Aggarwal, Senior Director of Product Management for Agent Evaluation & Responsible AI at ServiceNow, noted that on EVA-Bench it "sets a new Pareto frontier"(across both accuracy and customer experience dimensions, no other model outperforms it on both simultaneously).

Where It Is Already Being Used

Inside Tesla Vehicles: Handling millions of conversations daily. Vehicle owners can perform simple tasks via Connectors, such as checking calendars or replying to emails, or delegate larger tasks to Grok Bot, like ordering a coffee or booking dinner, entirely hands-free. Grok Bot's voice mode is also powered by Grok Voice 3.0.

Starlink Customer Service: Handling 20,000 multilingual calls daily with Grok Voice. Approximately 70% are resolved without requiring human intervention; among sales calls, 1 in every 4 results in an order placed before hanging up. An Agent built on Grok Voice 3.0 can execute the entire post-sales workflow: verifying customer identity, troubleshooting hardware failures, issuing billing credits, and arranging replacement shipments.

These figures are self-reported by xAI without disclosing the measurement timeframe or criteria for "resolved"; the Starlink section references "Grok Voice" as a whole, with only the individual Agent sentence explicitly specifying 3.0.

How to Use It

• General Users: Use the voice assistant inside Tesla vehicles, or use Grok Bot's voice mode.
• Developers: Integrate via xAI's Voice Agent API, or create and test voice Agents in the xAI Console; see integration instructions in the xAI Voice Documentation.
• Existing Deployments: Agents calling the grok-voice-latest model name have automatically switched to 3.0 with no code changes needed.

Pricing

Voice Agent API is 0.08 USD per minute (4.80 USD per hour), the same price as Grok Voice 2.0.

Artificial Analysis also compared another type of cost: the average spend per question when running Big Bench Audio. Grok Voice 3.0 is 0.025 USD, GPT-Live-1 (Astra, Medium) is 0.030 USD, and GPT-Realtime-2.1 (High) is 0.055 USD. On this test suite, Grok is about 17% cheaper than GPT-Live-1 and less than half the cost of GPT-Realtime-2.1. This figure is used for cross-model cost efficiency comparisons and cannot be directly converted into "how much a customer service call costs"—the duration and turn count of real calls differ from test questions.

GPT-Live-1's API bills separately for the "frontend audio layer" and backend model, which differs from Grok's flat per-minute pricing structure; metrics must be aligned before comparing prices:

Related ReadingOpenAI 开放 GPT‑Live‑1 API：可以将GPT的语音对话能力接入到任意应用中 每分钟 0.05 美元 (Chinese)Grok's primary rival this time: how it handles interruptions and noise, how it delegates tasks to the backend, and how tiered billing is calculated.→