SpaceXAI's Grok Voice 3.0 takes #1 on Artificial Analysis speech index

SpaceXAI's Grok Voice 3.0 has taken first place on the Artificial Analysis Speech-to-Speech Index with a composite score of 86.9, beating GPT-Live-1 (Astra, medium) at 83.2 and Gemini 3.8 Live Extended Thinking at 82.6. The realtime speech-to-speech model (a system that listens, reasons, calls tools and speaks in one session) ships at an unchanged $0.08 per minute. The announcement landed October 14, 2026.

The composite rank is the headline. The component scores tell a more mixed story, and SpaceXAI has left the model's size, architecture and context window undisclosed.

By James Calloway

What does the release actually change for developers?

Existing users woke up upgraded. The grok-voice-latest alias already points to grok-voice-3.0, so agents running the alias switched automatically. The endpoint, event format and per-minute price are unchanged from Grok Voice 2.0, which is now deprecated.

The API runs on SpaceXAI's hosted Voice Agent API only. There are no open weights, no local or quantized variants. The interface is compatible with the OpenAI Realtime API, so migration mostly means a new URL and key. Server-side voice activity detection handles turn-taking and barge-in, and reasoning is on by default at reasoning.effort: "high", settable to "none" for faster turns.

Tools run inside the session: collections search, web search, X search, remote MCP servers (a protocol for connecting models to external tools) and custom functions. Production teams should pin the versioned name grok-voice-3.0 rather than the alias.

The takeaway: the switch is free and automatic for alias users, but there is no way to run or audit the model outside SpaceXAI's infrastructure.

How strong are the benchmark numbers?

The AA Speech-to-Speech Index is an equal-weighted average of four components: Big Bench Audio reasoning, tau-Voice agentic performance, Speech Agent Arena preference and task success rate.

Grok Voice 3.0's component results are strong but not uniformly first:

• AA Index: 86.9, ranked #1.
• Big Bench Audio: 98.5%, ranked 4th. StepAudio 3 Realtime leads at 99.7%.
• Full Duplex Bench: 97.0%, ranked 4th. StepAudio 3 Realtime leads at 98.9%.
• Task success rate: 94.8%, up from 94.6% for Grok Voice 2.0.
• Speech Agent Arena Elo: 1070, the highest among the four models compared.
• tau-Voice: 71.6%, a large jump from 56.5%, but still behind GPT-Live-1 (Astra, medium) at 74.5%.

On ServiceNow's EVA-Bench, which scores enterprise voice agents on accuracy and experience, SpaceXAI reports a top accuracy score of 74%.

The composite #1 rests on breadth, not dominance: SpaceXAI is not the leader on agentic task resolution or raw speech reasoning, only on the average.

Is it fast and cheap?

Artificial Analysis measures 0.81 s time to first audio. GPT-Live-1 (Astra, medium) measures 1.34 s, so SpaceXAI's "40% lower latency" claim checks out against that rival. Grok Voice 2.0 was faster, though, at 0.70 s — a regression the announcement does not dwell on.

On cost, the comparisons depend on the metric:

• Cost per Big Bench Audio task: $0.025 for Grok Voice 3.0, versus $0.030 for GPT-Live-1 (Astra, medium) and $0.055 for GPT-Realtime-2.1 (High).
• Cost per hour of input audio: $4.80 for Grok Voice 3.0, versus $5.83 for GPT-Live-1 and $3.50 for Gemini 3.8 Live Extended Thinking.

List prices are not directly comparable. GPT-Live-1 costs $0.05 per minute but bills backend model and tool usage separately. Gemini 3.8 Live bills audio per minute plus text tokens ($0.005/min audio in, $0.018/min audio out).

Cheapest on one metric, not on another: Gemini undercuts SpaceXAI by $1.30 per input hour.

How does it stack up head-to-head?

Feature	Grok Voice 3.0	Grok Voice 2.0	GPT-Live-1 (Astra, medium)	Gemini 3.8 Live Ext. Thinking
Developer	SpaceXAI	SpaceXAI	OpenAI	Google
AA Index	86.9	81.3	83.2	82.6
tau-Voice	71.6%	56.5%	74.5%	68.6%
Task success rate	94.8%	94.6%	87.4%	89.1%
Arena Elo	1070	1011	1056	993
Time to first audio	0.81 s	0.70 s	1.34 s	1.35 s
Cost per hour input audio	$4.80	$4.80	$5.83	$3.50
List price	$0.08/min	$0.08/min	$0.05/min plus backend	$0.005/min in, $0.018/min out, plus text

SpaceXAI's blog lists frozen Arena Elo values of 1048 for GPT-Live-1 and 990 for Gemini, slightly below the live AA figures. Weights, parameter counts and context lengths are undisclosed for all four models. The bottom line in plain words: the lead is real on the composite, thin on the components.

What does deployment look like?

The API is OpenAI Realtime-compatible. Transports include WebSocket, WebRTC demo apps, and SIP (the telephony protocol) for phone and contact-center calls. Audio supports PCM at 8 to 48 kHz, G.711 μ-law and A-law, and Opus.

• Voices: 28 built-in, plus custom voices cloned from a reference clip of up to 120 s.
• Languages: 20 listed locales with automatic detection.
• Extensions: force_message for scripted lines, pronunciation replace maps, 30-minute session resumption.
• Compliance: SOC 2 Type II, HIPAA eligible with a BAA, and GDPR. SpaceXAI says audio is not stored or used for training.

That compliance posture targets the contact-center market, where latency and per-hour audio costs decide procurement — and where Google's cheaper hourly rate will compete directly.

What comes next?

Grok Voice 2.0 is deprecated, and the alias migration has already happened, so the practical question for developers is not whether to move but whether the 0.11 s latency regression and the tau-Voice gap against GPT-Live-1 matter for their workloads. Whether SpaceXAI holds the top composite slot will depend on the next moves from OpenAI and Google, whose models sit within 3.7 and 4.3 points respectively.