Grok Voice 3.0 leads the speech index. Support calls tell a more detailed story.

An upgrade that existing agents may already be using

SpaceX Super Intelligence, the company behind Grok, announced a speech-to-speech model: it takes spoken input and answers in audio, rather than simply reading a written script aloud. Its original launch post puts customer calls at the centre of the pitch.

The developer documentation confirms that grok-voice-latest now points to version 3.0. Integrations using that moving alias have already changed; those pinned to the deprecated 2.0 model need to update their model selection. The endpoint and session interface stay the same.

First overall does not mean first at everything

Independent evaluator Artificial Analysis tested the release through the Realtime API using its Rex voice. Grok scores 86.9 on the combined index, ahead of GPT-Live-1 (Astra, medium) at 83.2 and its own predecessor at 81.3.

That index gives equal weight to spoken reasoning, simulated support tasks, human preference and successful task-completing actions. Its methodology makes an important distinction: liking a conversation and completing its assigned task are separate measurements.

On the support-focused τ-Voice test, Grok rises from 56.5% to 71.6%, but GPT-Live-1 remains ahead at 74.5%. The test simulates callers and checks whether the agent leaves a database in the correct final state—for example, after changing a booking. Results average three trials where available; the Astra comparison currently uses one. These are benchmark outcomes, not live call-centre resolution rates.

The sub-second claim has a useful comparison

Artificial Analysis measures an average 0.81 seconds to first audio on Big Bench Audio, its spoken-reasoning test. That is quicker than Astra’s 1.34 seconds and Gemini’s high-thinking configuration at 1.35 seconds, but slower than Grok Voice 2.0 at 0.70 seconds. This release trades a little of its predecessor’s measured speed for stronger task performance.

The published tariff remains $0.08 per minute, or $4.80 an hour, with $0.004 per text input also listed. That is the voice API rate; it is not the complete cost of staffing a support operation. Artificial Analysis’s separate cost comparison runs a fixed audio workload and includes the model’s generated response. Its $4.80 result is below Astra’s $5.83 but above Gemini’s $3.50. Grok is not the cheapest option in every comparison.

ServiceNow finds an accuracy–experience trade-off

ServiceNow’s EVA-Bench adds a separate check across 213 scenarios covering customer service, IT support and HR. Its accuracy score requires a trial to clear thresholds for completing the task, speaking information faithfully and following the supplied context. Experience measures turn-taking, conciseness and conversational progress.

Grok leads the accuracy results below, while Gemini’s standard Live version leads on experience. The two columns answer different questions; they should not be merged into a single “calls resolved” percentage.

EVA uses simulated callers and some AI judges. Its limitations page warns that caller simulations miss aspects of real speech and that deployment choices can change timing. Multilingual expansion remains on the roadmap. These results do not establish how well an agent handles a Dutch or French customer queue.

Early users like the voice. Misunderstandings remain harder to diagnose.

KC, an AI commentator whose bio lists a FactoryAI ambassador role, says he tried the release in Grok Bot and liked its responsiveness and turn-taking. His post provides no recording, prompts or tool-action logs, so it is an encouraging impression rather than a repeatable comparison.

Quincy Lee, founder of Electric Era and a former SpaceX engineer, raises a different question in a reply to the launch discussion. He reports misinterpretations in Grok Bot voice mode and mentions his own A/B testing, without publishing the configurations or outputs.

Andrew Milich, whose profile lists SpaceXSI, asks Lee for a screenshot or request ID and says either the phone or model could be responsible. That exchange leaves the cause open. A consumer-app experience also includes the microphone, connection and application around the model.

For European teams, check the conversation and its location

The regional endpoint documentation says the global API can route requests between regions without guaranteeing the processing location. Its separate US endpoint excludes voice. The reviewed documentation therefore does not establish an EU residency guarantee for this realtime service.

The useful next test is specific: the languages customers speak, a corrected account number, an interruption during a tool action, and a request to reach a person. Track whether the underlying task ends correctly as well as how the conversation sounds. The independent evidence makes Grok Voice 3.0 a stronger candidate for that test; it does not make the test unnecessary.