SpaceXAI Grok Voice 3.0 Tops Artificial Analysis Speech-to-Speech Index With 86.9 Score

SpaceXAI released Grok Voice 3.0, a speech-to-speech model that now holds the top position on the Artificial Analysis Speech-to-Speech Index. According to MarkTechPost, the new model achieved a composite score of 86.9, surpassing previous leaders in the ranking. This release positions the company's flagship real-time voice technology as the highest-scoring entry currently tracked by the benchmarking organization.

The update introduces a model capable of listening, reasoning, calling tools, and speaking within a single session. SpaceXAI exposed the model as grok-voice-3.0 on the /v1/realtime WebSocket endpoint, with the grok-voice-latest alias automatically pointing to it. Developers using existing agents transitioned to the new version without manual intervention. The company did not publish parameter counts or architecture details for this release.

Performance Metrics and Benchmark Results

Grok Voice 3.0 leads the Artificial Analysis index, which averages Big Bench Audio reasoning, τ-Voice agentic performance, Speech Agent Arena preference, and task success rates. The model scored 86.9, outperforming GPT-Live-1 (Astra, medium) at 83.2 and Gemini 3.8 Live Extended Thinking at 82.6. Grok Voice 2.0 previously held a score of 81.3. Despite the overall lead, Grok Voice 3.0 ranked fourth in Big Bench Audio at 98.5%, behind StepAudio 3 Realtime's 99.7%. It also placed fourth in the Full Duplex Bench with 97.0%, compared to StepAudio 3 Realtime's 98.9%.

The model achieved a 94.8% task success rate, a slight increase from Grok Voice 2.0's 94.6%. In Speech Agent Arena Elo ratings, it reached 1070, the highest among the four models compared. However, it scored 71.6% on τ-Voice, which measures agentic performance, falling short of GPT-Live-1 (Astra, medium) at 74.5%. SpaceXAI also reported a top accuracy score of 74% on ServiceNow’s EVA-Bench, which evaluates enterprise voice agents.

Latency and Developer Integration

Artificial Analysis measured Grok Voice 3.0's time to first audio at 0.81 seconds. This figure supports SpaceXAI's claim of 40% lower latency compared to GPT-Live-1 (Astra, medium), which measured 1.34 seconds. However, Grok Voice 2.0 was faster, achieving 0.70 seconds. Reasoning is enabled by default with high effort but can be set to none for quicker turns. Server-side voice activity detection handles barge-in scenarios.

The model runs exclusively on SpaceXAI’s hosted Voice Agent API. There are no open weights, local, or quantized variants available for download. Tools run inside the session, including collections search, web search, X search, remote MCP servers, and custom functions. For developers, the endpoint, events, and pricing remain unchanged from the previous version at $0.08 per minute.

What should you do?

Security and technology professionals should monitor API integration points for any unexpected behavior when the grok-voice-latest alias switches to the new version. Since reasoning is on by default, teams should evaluate whether the increased processing time aligns with their latency requirements. Organizations using voice agents for enterprise tasks should review the τ-Voice agentic performance scores to determine if the model meets their specific resolution needs.

What to do and how to stay safe: SpaceXAI

• Review API integration documentation to ensure applications handle the automatic switch to grok-voice-3.0 via the latest alias.
• Test the new model's latency and reasoning capabilities in staging environments before deploying to production.
• Evaluate the $0.08 per minute cost against the performance gains to determine if the upgrade justifies the switch.
• Monitor tool execution logs for any changes in how the model handles remote MCP servers or custom functions.

Step-by-step guide: Mobile Device Security: How to Protect Data Beyond Passwords

General security guidance from the Threat Times newsroom. It is not confirmed advice from the organisations named in this story.

Frequently asked questions

What score did Grok Voice 3.0 achieve on the Artificial Analysis index?

Grok Voice 3.0 achieved a score of 86.9 on the Artificial Analysis Speech-to-Speech Index, ranking first.

Is Grok Voice 3.0 available for local deployment?

No, Grok Voice 3.0 runs only on SpaceXAI’s hosted Voice Agent API with no open weights or local variants.

How does the pricing compare to Grok Voice 2.0?

The price remains unchanged at $0.08 per minute, with the same endpoint and events structure.