xAI launches Grok Voice 3.0, saying it ranks Speech-to-Speech Index first
BIBIBI
AT A GLANCE
xAI announces the launch of the Speech-to-Speech model Grok Voice 3.0, saying it ranks first on the benchmark, with latency below 1 seconds and pricing of 0.08 USD。
Article
xAI launches Grok Voice 3.0, ranking first in the Speech-to-Speech benchmark
Beating AI news brief: xAI announced the launch of the Speech-to-Speech model Grok Voice 3.0, saying it ranked first with a score of Speech-to-Speech Index in an independent Artificial Analysis test of the 86.9 benchmark, higher than the 2.0 score of Grok Voice 81.3、GPT-Live-1 and the 83.2 score of Gemini 3.8 Live; it also said end-to-end latency was below 82.6 seconds. The model is now available through 1, priced at Voice Agent API per minute, unchanged from the previous generation; agents using grok-voice-latest will switch automatically.
According to the company, Grok Voice handles millions of conversations daily at Tesla and SpaceX. Starlink customer service centers reportedly answer about 0.08 USD calls per day, with about 20000 resolved without human intervention; about one-quarter of sales calls result in a purchase before the call ends. ServiceNow 70%-Bench accuracy is EVA, while τ-Voice is 74%, reportedly comparable to 71.6%but with latency approximately GPT-Live-1 lower.
Original link 40%https://m.theblockbeats.info/flash/371851
Key points
01
xAI announces the launch of the Speech-to-Speech model Grok Voice 3.0。
02
xAI says Grok Voice 3.0 ranked first in the independent Artificial Analysis test of the Speech-to-Speech Index with a score of 86.9.
03
The comparative scores listed in the article are: Grok Voice 2.0 scored 81.3、GPT-Live-1 scored 83.2、Gemini 3.8 Live scored 82.6。
04
xAI says the model's end-to-end latency is below 1 seconds.
05
The model is now available through Voice Agent API, priced at 0.08 USD per minute, unchanged from the previous generation.
06
Agents using grok-voice-latest will switch automatically.
07
According to the company, Grok Voice handles millions of conversations daily at Tesla and SpaceX.
08
According to the company, Starlink customer service centers answer about 20000 calls per day, with about 70%resolved without human intervention; about one-quarter of sales calls result in a purchase before the call ends.
09
The article says ServiceNow EVA-Bench accuracy is 74%, while τ-Voice is 71.6%, comparable to GPT-Live-1 but with latency approximately lower by 40%。
AI-assisted interpretation
The following is analysis, separate from reported facts. Verify important claims independently.
This is a news report about the release of a voice model. “Speech-to-Speech” means that the system directly receives speech and responds with speech. The report says the new model is already available through Voice Agent API and provides pricing, latency, benchmark results, and business data from customer service scenarios; several of these data points are claims by xAI or the company.
Why it matters to readers
If the reported results, latency, and customer service data can be reproduced, Grok Voice 3.0 may suit agent and customer service scenarios requiring rapid voice responses. The price of 0.08 USD per minute and the automatic switching arrangement will also directly affect existing API users.
Beginners should focus on three indicators: the per-minute price, actual response latency, and whether third-party testing can reproduce the official results. Benchmark rankings do not necessarily correspond to the experience in all real-world scenarios.
Risks and unknowns
The evidence does not include an official xAI announcement or Voice Agent API documentation, so the launch, pricing, and automatic switching arrangements cannot be independently verified from the materials currently available.
The evidence does not provide the original Artificial Analysis ranking or testing methodology, so the claim of a 86.9 score and first place cannot be independently verified.
The conversation volume, automated resolution rate, and purchase conversion data for Tesla, SpaceX, and Starlink are all company claims, with no third-party verification in the evidence.
Related Developments
Loading event timeline…
ServiceNow EVA-Bench and τ-Voice results and the latency comparison lack testing conditions and original reports.
The evidence does not specify the measurement environment or scope of applicability for the end-to-end latency below 1 seconds.
Related concepts
Speech-to-Speech
This term is not in the glossary yet. Browse related concepts in the glossary.
