xAI opens Grok Voice 3.0 in the Voice Agent API

What did SpaceXSI announce on 14 October?

SpaceXSI published an announcement titled “Introducing Grok Voice 3.0,” describing the release as its newest speech-to-speech model for the Voice Agent API. The post states that the model is available today at the same price as its predecessor and that agents already using the grok-voice-latest alias were switched over automatically, with no API shape change required.

The company frames Grok Voice as already powering millions of conversations a day across Tesla and SpaceX products, and it positions 3.0 as improving customer-support resolution, latency under one second, and cost relative to other frontier voice models. Those comparative claims rest on cited Artificial Analysis and ServiceNow measurements rather than on a public, reproducible eval package hosted by SpaceXSI itself.

For readers comparing voice stacks, our earlier coverage of Fish Audio’s Drama 3 open-weight TTS and Mistral’s Voxtral realtime Arabic release shows how other vendors split speech understanding, speech generation, and realtime duplex into different products. Grok Voice 3.0 is presented as a single speech-to-speech agent path.

Which model IDs and prices should developers use?

Developer documentation on docs.x.ai lists grok-voice-3.0 as the flagship speech-to-speech model and maps grok-voice-latest to that release. Grok Voice 2.0 is marked deprecated: sessions that still pin grok-voice-2.0 should set model to grok-voice-3.0 or grok-voice-latest. The announcement and docs both say the endpoint, session parameters, and event protocol are unchanged from 2.0.

List pricing on the announcement is $0.08 per minute, or $4.80 per hour, matching Grok Voice 2.0. That per-minute rate is the figure SpaceXSI publishes for the Voice Agent API; it is not a guaranteed cost-per-task once tool calls, reasoning effort, or telephony transport are layered on. Teams that need a stable bill should pin the versioned model id rather than the moving latest alias.

The realtime path is a WebSocket Speech to Speech API at wss://api.x.ai/v1/realtime with model passed as a query parameter. Docs describe OpenAI Realtime–compatible client patterns with a base URL and key swap, plus SpaceXSI-specific extensions such as force_message scripted TTS lines and session resumption. Authentication can use an API key server-side or ephemeral tokens for browsers and mobile clients.

• Pin grok-voice-3.0 in production if you need a fixed model; use grok-voice-latest only when you accept automatic upgrades.
• Budget $0.08/min for the voice path as published; measure tool and reasoning add-ons separately.
• Prefer ephemeral tokens for client-side apps so the long-lived API key never ships in the browser.

What do the cited benchmarks actually measure?

The announcement’s main table attributes an overall Artificial Analysis Speech-to-Speech Index of 86.9 to Grok Voice 3.0, ahead of Grok Voice 2.0 at 81.3, GPT-Live-1 (Astra, Medium) at 83.2, and Gemini 3.8 Live (Extended Thinking, High) at 82.6. Sub-scores include Speech Reasoning Big Bench Audio at 98.5%, Speech Agent Arena Elo at 1070, Task Success Rate at 94.8%, and Full Duplex Bench at 97.0%. The table is labeled as sourced from Artificial Analysis, October 13, 2026.

Separately, SpaceXSI says τ-Voice—Artificial Analysis’s support-call suite with accents, noise, and interruptions—scores 71.6% for Grok Voice 3.0, “comparable with GPT-Live-1 (Astra)” while claiming about 40% lower latency. A cost plot on the page claims a lower average cost per task on Artificial Analysis’s Big Bench Audio run than GPT-Live-1 (Astra, Medium) and GPT-Realtime-2.1. AI Lookout did not download Artificial Analysis’s raw runs or verify latency clocks.

On ServiceNow’s EVA-Bench, the announcement quotes a 74% accuracy score and a joint accuracy-plus-experience lead, with a named quote from Nitin Aggarwal of ServiceNow. EVA-Bench is described there as an open-source enterprise voice-agent benchmark spanning more than 200 scenarios. Again, the score is reported by the vendor page; it is not an AI Lookout replication.

Where is SpaceXSI already running voice agents?

Operationally, the announcement says Grok Voice 3.0 handles millions of conversations a day in Tesla vehicles, covering Connectors for short tasks such as calendar checks and Grok Bot for longer hands-free tasks such as ordering or booking. It also says the same model powers Grok Bot’s voice mode. Those product surfaces are SpaceXSI’s own deployments; outside developers get the API rather than Tesla’s in-car stack.

For Starlink call centers, SpaceXSI writes that Grok Voice answers about 20,000 customer calls every day across multiple languages, that roughly 70% are resolved with no human in the loop, and that one in four sales calls ends in a purchase before the caller hangs up. The post describes a single agent that can verify customers, troubleshoot hardware, issue credits, and ship replacement equipment. Those percentages are company operational claims, not audited third-party metrics.

Voice-agent demos elsewhere in the industry often separate controlled arenas from production contact-center traffic. Our note on HeyGen Voice’s controlled arena results is a reminder that leaderboard settings and live call mixes rarely match one-for-one.

How should teams evaluate Grok Voice 3.0 before swapping a stack?

Treat the Artificial Analysis and EVA-Bench numbers as directional, then run a narrow eval on your own call types: accents you actually hear, the languages you staff, the tools your agent must call, and the telephony codecs you already use. Docs list PCM, μ-law, A-law, and Opus options, plus language hints and keyterms for transcription bias—features that matter more on narrowband phone audio than on clean headset demos.

Because grok-voice-latest already points at 3.0, staging environments that used the alias may already be on the new model. Pin 2.0 only if you still need a rollback path, and schedule a cutover test that compares interruption handling, tool-call latency after audio finishes playing, and pronunciation replacements for brand names. SpaceXSI’s docs warn that sending response.create before playback finishes can overlap audio after tool calls.

Finally, separate product marketing from contractual commitments. The announcement’s “world’s most intelligent” framing is vendor language. What you can verify today is the model id, the published per-minute price, the alias switch, and whether your own success metrics improve after a controlled canary.

• Canary a versioned grok-voice-3.0 deployment against a recorded slice of real calls.
• Measure barge-in, tool-call turn gaps, and escalation rates—not only leaderboard Elo.
• Keep a pinned prior model until your own metrics clear a pre-agreed bar.

What remains unclear or out of scope?

The announcement does not publish a full safety card, training-data summary, or independent red-team package for Grok Voice 3.0 in the same post. It also does not break out pricing for tool calls, reasoning.effort high versus none, or SIP telephony beyond pointing readers at separate docs. Availability claims are for the Voice Agent API; regional restrictions, enterprise contracting, and rate limits are not enumerated on the launch page.

AI Lookout has not placed live calls through Starlink support, has not instrumented Tesla cabin audio, and has not re-scored Artificial Analysis or EVA-Bench. Readers should treat deployment statistics and third-party tables as attributed claims until they can reproduce them or obtain contractual SLAs.

If SpaceXSI later publishes a dated changelog with different prices, model ids, or benchmark revisions, that would be material update territory for this article rather than a separate launch story.