Grok Voice 3.0 ranks first on benchmarks, adds tool-calling

xAI released Grok Voice 3.0, a speech-to-speech voice model that ranks first on independent benchmarks and introduces tool-calling capabilities for developers building voice agents. The company announced the launch on October 14, 2026, with the model available immediately through the Voice Agent API at the same price as its predecessor.

What changed

Grok Voice 3.0 is positioned as an upgraded voice model that listens, reasons, calls tools, and speaks. According to Artificial Analysis independent testing, the model now ranks first on the Speech-to-Speech Index with a score of 86.9, compared to 83.2 for GPT-Live-1 (Astra, medium) and 82.6 for Gemini 3.8 Live Extended Thinking. Grok Voice 2.0 scored 81.3 on the same index.

The new version adds tool-calling capabilities inside the voice session. Developers can now configure web search, X search, collections search, remote MCP servers, and custom functions that execute during conversations. Reasoning runs at “high” effort by default but can be set to “none” for faster response times. Voice activity detection on the server side handles interruptions automatically.

Technically, Grok Voice 3.0 runs as a proprietary API-only service on xAI’s hosted Voice Agent infrastructure. The company has not published parameter counts or architecture details, and there are no open weights or local variants. The model is exposed as grok-voice-3.0 on the /v1/realtime WebSocket endpoint, and existing agents using the grok-voice-latest alias switched automatically without requiring code changes.

Pricing and deployment options remained unchanged: the model costs $0.08 per minute, identical to Grok Voice 2.0. xAI supports WebSocket, WebRTC, and SIP transports for phone and contact-center integration. Audio codec support includes PCM at 8 to 48 kHz, G.711 mu-law and A-law, and Opus. The model offers 28 built-in voices plus custom voice cloning from up to 120 seconds of reference audio. Language support covers 20 locales with automatic language detection.

Why it matters

The ranking shift indicates progress in voice reasoning and task completion for real-world applications. xAI reports that Grok Voice 3.0 powers millions of daily conversations at Tesla and Starlink, with Starlink call centers reporting that roughly 70% of customer calls resolve without human intervention. This scale demonstrates that tool-integrated voice agents are moving beyond experimental use into production workloads that handle complex, multi-step tasks like hardware troubleshooting, credit issuance, and equipment shipping.

For developers, the tool-calling addition removes friction from voice agent workflows. Previously, voice models could reason but could not directly execute external functions. The integrated approach means developers can now build agents that answer questions by searching the web during a conversation or that complete tasks like verifying customer accounts and issuing credits without leaving the voice session. This matters because it reduces latency and context switching compared to orchestrating separate APIs.

The latency profile also reshapes what is possible for time-sensitive applications. Grok Voice 3.0 achieves 0.81 seconds time to first audio, 40% lower latency than GPT-Live-1 (Astra, medium). This matters for customer support calls where silence harms engagement and for in-vehicle voice assistants where responsiveness influences perception.

However, the model does not lead uniformly across all benchmarks. On tau-Voice, which tests handling of diverse accents, background noise, and interruptions, Grok Voice 3.0 scores 71.6%, below GPT-Live-1 (Astra, medium) at 74.5%. On Big Bench Audio reasoning, it scores 98.5% but ranks fourth, behind StepAudio 3 Realtime at 99.7%. This signals that while the composite score is highest, specialists and enterprise buyers should verify task-specific performance before switching.

Similar to how multi-agent AI teams require careful testing, voice agent deployments benefit from benchmarking against real workloads. xAI’s report of 74% accuracy on ServiceNow’s EVA-Bench is strong, but enterprise teams should validate accuracy and customer experience on their own call types before deploying.

What to test

Before adopting Grok Voice 3.0, teams should verify several vendor claims and test edge cases:

1. Tool-calling latency under load: Test whether the 0.81-second time to first audio holds when tools are invoked. Web search and MCP server calls add latency; measure real-world performance with your typical tool mix.

2. Tau-Voice performance on your customer base: If your users include diverse accents or noisy environments, compare Grok Voice 3.0 to GPT-Live-1 on sample calls. The 71.6% score is lower than the competing model.

3. Cost per task in your workflow: xAI cites $0.025 per Big Bench Audio task, but your task mix may differ. Measure actual usage patterns to understand cost implications.

4. Custom voice cloning quality: Test the voice cloning feature on your reference audio. The 120-second limit may not capture all nuances of accented or highly distinctive voices.

5. Session resumption and compliance: Verify that HIPAA eligibility and GDPR compliance match your requirements. Confirm that audio is not retained or logged.

6. Interrupt handling and Full Duplex performance: Although Grok Voice 3.0 scores 97% on Full Duplex Bench, test natural pause and turn-taking behavior on your call types.

Common questions

Can Grok Voice 3.0 call tools in voice mode?

Yes. Grok Voice 3.0 now supports tool integration inside the voice session, including web search, X search, collections search, remote MCP servers, and custom functions. Tools execute during the conversation without leaving the voice API.

How does Grok Voice 3.0 compare on cost per minute?

Grok Voice 3.0 costs $0.08 per minute, the same as Grok Voice 2.0. According to xAI’s analysis of Big Bench Audio task cost, it runs at $0.025 per task, below GPT-Live-1 (Astra, medium) at $0.030 and GPT-Realtime-2.1 (High) at $0.055. Other models bill separately for backend model and tool usage, so direct hourly comparison is not straightforward.

Is Grok Voice 3.0 available on all platforms?

Grok Voice 3.0 runs API-only on xAI’s hosted infrastructure. There are no open weights, no local deployment options, and no quantized variants for edge devices.

The conclusion

Grok Voice 3.0 represents progress in voice agent capability. A top composite benchmark score, sub-second latency, and integrated tool-calling move the bar for production applications. The shift from pure reasoning to reasoning-plus-action workflows reduces deployment complexity for teams building customer support or in-vehicle voice interfaces.

However, the benchmark leadership is not universal. Lower tau-Voice performance and variable Big Bench Audio ranking mean specialists should validate on their own use cases. The API-only availability also locks teams into xAI’s infrastructure, which removes deployment flexibility compared to open-weight alternatives.

What to watch next: whether the tool-calling integration becomes standard across other frontier voice models, and whether xAI publishes parameter counts or architecture details that would let researchers and competitors understand the reasoning gains. The scale at Tesla and Starlink suggests the model is production-ready, but independent voice agent deployments in regulated industries should test compliance and audit trails carefully before switching from incumbents.

Related reading

• LFM2.5-24B-A2B: Fast on-device reasoning agents
• RTX Spark PC AI launches Oct 16 with local agents
• QwenGyre elastic RL framework trains long-horizon agents