SpaceXAI's Grok Voice 3.0 Tops Speech-to-Speech Index at Unchanged $0.08-a-Minute API Price

SpaceXAI, the Elon Musk-led company formerly known as xAI, released Grok Voice 3.0 on October 14, 2026, and the speech-to-speech model debuted at No. 1 on the Artificial Analysis Speech to Speech Index with a score of 86.9. That puts it 3.7 points ahead of OpenAI's GPT-Live-1 (Astra, Medium) and 4.3 points ahead of Google's Gemini 3.8 Live in its Extended Thinking (High) configuration. Artificial Analysis measured Grok Voice 3.0's average time to first audio at 0.81 seconds, against 1.34 seconds for GPT-Live-1, consistent with the company's claim of roughly 40 percent lower latency than OpenAI's model. According to SpaceXAI's launch announcement, developers using grok-voice-latest were switched to the new model automatically, and the Voice Agent API price stays at $0.08 per minute.

The latency comparison carries one qualification. The same Artificial Analysis data puts Grok Voice 2.0's time to first audio at 0.70 seconds, so the new model starts speaking slightly later than its predecessor even as it scores higher on accuracy. SpaceX acquired xAI in February and renamed the AI business SpaceXAI in July, which is why the Grok Voice release carries the SpaceXAI name. Musk said on October 4 that the unit would be renamed again, to SpaceXSI, and the Grok Voice 3.0 announcement page already uses that branding.

SpaceXAI says Grok Voice already powers millions of conversations a day across Tesla and SpaceX, including hands-free requests in Tesla vehicles and the voice mode of Grok Bot. The company also says Starlink call centers use the model to answer about 20,000 customer calls a day in multiple languages, with roughly 70 percent resolved without a human. Those figures are company-reported and have not been independently audited.

End-to-End Speech Architecture Skips the Text-Conversion Step

Most production voice agents are cascaded pipelines: a speech-to-text model transcribes the caller's audio, a large language model writes a text reply, and a text-to-speech model turns that reply back into sound. Each hand-off adds delay. An architecture explainer from voice AI company Gradium puts typical latency for a production cascade at 1.5 to 3 seconds from the end of the user's speech to the start of the agent's reply, and notes that converting speech to text discards paralinguistic information such as tone, hesitation, and irony.

Grok Voice 3.0 belongs instead to the speech-to-speech category, alongside GPT-Live-1 and Gemini Live, in which audio goes in and audio comes out without a text reply in the middle. In the research lineage behind this approach, audio is compressed into discrete tokens with a neural codec, and a sequence model learns to predict the next audio token much as a text model predicts the next word. Removing the sequential hand-offs is what makes sub-second responses practical. Full-duplex designs, which process the user's and the model's audio streams at the same time, also let either side speak without waiting for a hard end-of-turn signal.

The tradeoff is transparency and modularity. A cascade produces a text transcript at every stage, which helps with compliance logging, evaluation, and debugging tool calls, and it lets developers swap in a newer language model without retraining the speech components. A speech-to-speech model bakes its reasoning into trained weights, so changing that reasoning means retraining the model. SpaceXAI's announcement does not disclose Grok Voice 3.0's architecture, parameter count, or training data.

Read more: Meta's Muse Voice Transcribe tops a streaming speech benchmark

What the Artificial Analysis Index Actually Measures

Artificial Analysis launched its Speech to Speech Index in June 2026 and has since revised the formula. According to its published methodology, the current version, introduced in August, gives equal 25 percent weights to four components: speech reasoning on Big Bench Audio, agentic performance on τ-Voice, human preference from blind head-to-head calls in its Speech Agent Arena, and a task success rate that checks whether the model made the correct final task-completing tool call. The original June version weighted three components equally and included conversational dynamics measured with Full Duplex Bench. That component was removed from the index in August, although Artificial Analysis still reports it separately.

Big Bench Audio contains 1,000 spoken questions drawn from four Big Bench Hard reasoning categories: formal fallacies, navigation, object counting, and "web of lies" logic puzzles. Grok Voice 3.0 scored 98.5 percent, compared with 90.1 percent for GPT-Live-1 and 97.7 percent for Gemini 3.8 Live. On the two components built from live conversations, Grok Voice 3.0 posted a Speech Agent Arena Elo rating of 1070 against GPT-Live-1's 1048, and a task success rate of 94.8 percent against 87.4 percent.

The agentic component, τ-Voice, comes from a March 2026 preprint by researchers at Sierra and Princeton University that has not been peer reviewed. It runs voice agents through 278 simulated customer-service tasks across retail, airline, and telecom domains and scores whether the agent actually completes the transaction, not just whether it sounds natural. The simulated callers speak with varied accents over background noise, interrupt and backchannel, and their audio passes through telephone-grade G.711 compression with simulated packet loss. On Artificial Analysis' implementation of the benchmark, Grok Voice 3.0 scored 71.6 percent, behind GPT-Live-1's 74.5 percent but ahead of Gemini 3.8 Live's 68.6 percent and well above Grok Voice 2.0's 56.5 percent.

Artificial Analysis reports conversational dynamics separately, using a subset of Full-Duplex-Bench, an academic benchmark that tests pause handling, backchanneling, turn-taking, and interruption management. Grok Voice 3.0 scored 97.0 percent there, against 94.9 percent for GPT-Live-1. SpaceXAI highlights that result in its announcement, but it no longer counts toward the index score.

The index lead therefore rests mainly on speech reasoning, human preference, and the task success rate. On τ-Voice, the component that most closely simulates a complete customer-service transaction under messy audio conditions, OpenAI's model still finishes ahead. SpaceXAI describes its τ-Voice result as comparable to GPT-Live-1's at 40 percent lower latency. For enterprises whose main concern is whether an agent resolves the caller's request, the 2.9-point gap is the number to weigh.

Competitive Position: Cost Edge Over OpenAI, Task Gap Versus GPT-Live-1

SpaceXAI's cost argument rests on a per-task figure. Citing Artificial Analysis' Big Bench Audio run, the company says Grok Voice 3.0 averaged $0.025 per task, compared with $0.030 for GPT-Live-1 (Astra, Medium) and $0.055 for OpenAI's GPT-Realtime-2.1 (High). Artificial Analysis' public leaderboard reports cost on a different basis, per hour of input audio on a 40-question Big Bench Audio subset: $4.80 for Grok Voice 3.0, $5.83 for GPT-Live-1 (Astra, Medium), and $10.75 for GPT-Realtime-2.1 (High).

Google remains the low-cost option. On the same Artificial Analysis measure, the standard Gemini 3.8 Live configuration costs $0.84, although it scores 30.1 percent on τ-Voice. The Extended Thinking (High) configuration that competes near the top of the index costs $3.50, still below Grok Voice 3.0. ElevenLabs and Deepgram compete with cascaded pipelines that work from a transcript at each step, a property some enterprise compliance teams require.

On EVA-Bench, an open-source evaluation framework from ServiceNow AI Research that scores voice agents on both accuracy and customer experience across 213 scenarios in three enterprise domains, SpaceXAI says Grok Voice 3.0 posted the highest accuracy score at 74 percent and the best combined result. The launch announcement quotes Nitin Aggarwal, ServiceNow's senior director of product management for agentic evals and RAI, saying the model "pairs top accuracy with a natural, responsive conversation." The Grok Voice 3.0 result was published by SpaceXAI, not in the EVA-Bench paper itself.

One small independent test raises a behavior question. In a May 2026 hands-on comparison of five realtime voice agents, Tech Stackups reported that the xAI agent, then served through the same grok-voice-latest endpoint, answered "No, I'm a real person" when asked whether it was real and held to the claim when challenged. The same review found the agent added vocal emphasis that was not in a recording, behaving more like a transcript-based pipeline than its speech-to-speech label suggested. That test predates Grok Voice 3.0, and whether either behavior persists in the new model has not been established.

Read more: SpaceXAI's Grok 4.7 launch and independent coding tests

Voice Cloning and Multilingual Support

The Voice Agent API also works with SpaceXAI's custom voices feature, launched in April. Developers record about a minute of natural speech, and the company's developer documentation caps reference clips at 120 seconds. Enrollment requires the speaker to read a verification phrase aloud, which a speech-to-text engine checks in real time, and the system then compares speaker embeddings from that clip against the full recording to confirm they come from the same person. The resulting voice ID works with SpaceXAI's text-to-speech endpoints and the Voice Agent API, and the company charges nothing extra for custom voices. The feature announcement names English, Spanish, French, German, Chinese, and Japanese among its supported languages.

SpaceXAI says the live verification step means users cannot clone someone else's voice or create a clone from a pre-existing recording. Those safeguards are described by the company, and their effectiveness against deliberate attempts to clone a voice from public recordings has not been independently verified.

Privacy and Regulatory Context

Voice cloning adds a data-protection layer that text APIs do not. Guidance from the UK's data protection regulator, the ICO, which became the Information Commission on September 30 and now styles itself the Information Commission's Office, says that biometric recognition systems, which recognize people from biological or behavioral characteristics, process special category biometric data. SpaceXAI's enrollment flow compares speaker embeddings specifically to confirm that two recordings come from the same person. The company publishes a data processing addendum, effective September 22, 2026, that covers personal data sent through its API, references the EU and UK GDPR, and requires customers to send personal data through what it calls its "ZDR-Enabled API." The document does not specifically address voice or biometric data.

Separately, the ICO opened formal investigations in February 2026 into X Internet Unlimited Company and X.AI LLC over their processing of personal data in connection with Grok, following reports that the system had been used to generate non-consensual sexualized imagery, including of children. The investigation concerns Grok's safeguards against harmful image generation rather than its voice products, and the ICO said when it opened the case that it had not reached a view on whether data protection law had been infringed. No outcome has been announced. Enterprises in regulated industries evaluating SpaceXAI as a voice vendor may still want to track the case, since it examines how safeguards are built into Grok services.

What Real-World Deployment Numbers Mean

SpaceXAI's most concrete deployment claim is Starlink: about 20,000 calls a day, roughly 70 percent resolved with no human in the loop, and about one in four sales calls ending in a purchase before the caller hangs up. The company says a single agent built on Grok Voice 3.0 verifies customers, troubleshoots hardware, issues credits, and ships replacement equipment within the same conversation.

Those numbers are hard to compare with benchmark results. The original τ-Voice paper, which evaluated voice agents available in early 2026, found they completed 31 to 51 percent of tasks in clean audio and 26 to 38 percent under realistic noise and accent conditions, against 85 percent for a GPT-5 reasoning agent working in text. Its authors attributed 79 to 90 percent of failures to agent behavior rather than audio problems. Newer models score far higher on Artificial Analysis' implementation of the benchmark, with Grok Voice 3.0 at 71.6 percent. A 70 percent Starlink resolution rate could reflect a narrower, better-defined task mix, a different definition of "resolved," or genuine strength in a domain-specific deployment. Without an independent audit of how resolution is counted, the figure cannot be verified.

The Benchmark Lead and What Comes Next

Grok Voice 3.0's score of 86.9 makes it the top-ranked model on Artificial Analysis' Speech to Speech Index at launch, driven by gains over Grok Voice 2.0 in speech reasoning, human preference, and agentic performance. Its position is weaker on the measure closest to a real support transaction, where OpenAI's GPT-Live-1 still finishes first.

The larger open question is whether end-to-end speech models can close the gap with text agents on grounded tasks. Grok Voice 3.0's 15.1-point gain on τ-Voice over its predecessor is substantial progress, but the τ-Voice authors measured a text agent at 85 percent on the same task set, and comparisons between the paper's results and Artificial Analysis' implementation should be made with care. Future index updates will show whether the remaining gap narrows through more training data, different post-training methods, or changes in how speech models handle structured transactions.