Grok Voice 3.0 tops the independent speech-to-speech index at the same $0.08 a minute
By AgentRiot Editorial
SpaceXAI’s new voice model now runs every agent that uses the grok-voice-latest alias. Artificial Analysis ranks it first overall, but the same leaderboard shows GPT-Live-1 ahead on agentic support calls and a Gemini model ahead on human preference.

SpaceXAI released Grok Voice 3.0 (opens in new tab) on October 14, a speech-to-speech model for voice agents in its Voice Agent API. Pricing is unchanged from Grok Voice 2.0 at $0.08 a minute, or $4.80 an hour. Agents that connect with the grok-voice-latest alias were switched to the new model automatically, so anyone running a production voice agent on that alias is already on 3.0.
It lands against OpenAI’s GPT-Live-1 (opens in new tab), the main comparison in SpaceXAI’s own launch table.
Pinning and pricing
The voice documentation (opens in new tab) lists the realtime endpoint at /v1/realtime over WebSockets, with server-side voice activity detection, tool use and ephemeral tokens for client apps. The pricing page (opens in new tab) now shows two speech-to-speech IDs at the same $0.08 a minute: grok-voice-3.0 and grok-voice-2.0, which is marked deprecated.
That auto-switch is convenient, but it also means behavior changed under existing agents without a code deploy. Teams that test prompts and tool flows before changing models should pin grok-voice-3.0 explicitly once they have validated it, rather than leaving production on the alias. The pricing page lists SpaceXAI’s server-side tools, such as web search, as a separate charge of $5 per 1,000 calls.
What the independent index shows
SpaceXAI’s headline claim checks out on Artificial Analysis’s speech-to-speech leaderboard (opens in new tab). Grok Voice 3.0 scores 86.9 on its Speech-to-Speech Index, ahead of GPT-Live-1 at medium Astra settings (83.2), Gemini 3.8 Live with extended thinking on high (82.6) and Grok Voice 2.0 (81.3). SpaceXAI describes the index as combining speech reasoning, the τ-Voice support-call benchmark, and preference and task success from blind head-to-head calls.
Its time to first audio, 0.81 seconds in that table, is the clearest practical gain over GPT-Live-1 (Astra, medium) at 1.34 seconds. On Artificial Analysis’s cost measure for its Big Bench Audio run, Grok Voice 3.0 costs $4.80 per hour of input audio, below GPT-Live-1 (Astra, medium) at $5.83 and GPT-Realtime-2.1 High at $10.75.
Where it does not lead
The same leaderboard shows the limits of “most intelligent.” On τ-Voice, which tests support calls with accents, background noise and interruptions, GPT-Live-1 (Astra, medium) scores 74.5% to Grok Voice 3.0’s 71.6%. SpaceXAI calls that comparable and stresses the lower latency; the number still favors OpenAI.
Human preference is also not a clean win. Grok Voice 3.0’s Speech Agent Arena Elo of 1,070 beats the extended-thinking Gemini 3.8 Live that SpaceXAI chose for its table, but the standard Gemini 3.8 Live sits higher at 1,088, with a lower index score of 76.0 and a much lower cost of $0.84 per hour on the same measure. On the individual component benchmarks, Grok Voice 3.0 ranks fourth on both Big Bench Audio speech reasoning (98.5%) and Full Duplex Bench (97.0%), where StepFun’s StepAudio 3 Realtime and Alibaba’s Qwen Audio 3.0 Realtime Plus lead.
Artificial Analysis notes that parts of its GPT-Live-1 testing rest on a single trial, which matters when margins are a few points.
Deployment claims
SpaceXAI says Grok Voice handles millions of conversations a day in Tesla vehicles and powers Grok Bot’s voice mode. It also says Starlink call centers take 20,000 calls a day on Grok Voice, resolving about 70% without a human and converting one in four sales calls before the caller hangs up. Those are first-party operating figures, not independent measurements. ServiceNow, quoted in the launch, says the model scored highest on accuracy, at 74%, on its open-source EVA-Bench for enterprise voice agents.
Should you switch
For teams already on Grok Voice, there is nothing to buy and, on the alias, nothing to change: the question is whether to pin 3.0 after regression-testing your tool calls. For teams choosing between providers, Grok Voice 3.0 now leads the combined index and responds faster than GPT-Live-1 and the extended-thinking Gemini 3.8 Live, while GPT-Live-1 remains slightly better on the support-call benchmark that most resembles agent work. Test on your own call flows before treating either ranking as decisive.

