Grok Voice 3.0 is the speech-to-speech model xAI shipped on October 14, 2026, for developers building voice agents. It replaces Grok Voice 2.0 as the default behind the Voice Agent API. The model targets phone support, contact centers, and in-app voice assistants. xAI reports a first-place score on Artificial Analysis’s Speech-to-Speech Index. For teams already running Grok Voice 2.0, the upgrade arrived automatically through a shared model alias. That silently changed call behavior for anyone who never pinned a specific version.
What Is Grok Voice 3.0?
Grok Voice 3.0 is xAI’s third-generation speech-to-speech model for real-time voice agents. Unlike a pipeline that chains speech-to-text, a language model, and text-to-speech, this model processes and responds in audio directly. xAI built it for phone support lines, contact centers, and in-app assistants that need natural, low-latency conversation. The company released it through the Voice Agent API on October 14, 2026. That model carries the ID grok-voice-3.0.
xAI is now part of SpaceXAI, the combined entity Elon Musk formed when SpaceX absorbed xAI in July 2026. The Grok brand and developer documentation still live at the x.ai domain. This piece uses “xAI” for the team that ships Grok Voice, for clarity. That corporate detail matters less to engineers than what changed in the model itself.
How Grok Voice 3.0 Works
Developers reach Grok Voice 3.0 through a WebSocket endpoint at wss://api.x.ai/v1/realtime. The model ID goes in the query string. That API matches the shape of OpenAI’s Realtime API. Most teams only change a URL and an API key to connect. Voice agents extend the same generative AI shift already reshaping how teams build software. That shift now moves from writing code to handling live customer calls. Grok Voice 3.0 also supports WebRTC demo apps and SIP, which covers phone and contact-center calls directly.
The model handles 20 languages with automatic detection, so it can switch mid-call without a manual language flag. It ships 28 built-in voices. Teams can also clone a custom voice from a reference clip up to 120 seconds long. Server-side voice activity detection manages interruptions, letting a caller talk over the agent naturally.
Reasoning runs on by default at a high effort setting, which helps the model plan multi-step answers before speaking. Teams that need faster, simpler replies can set reasoning.effort to “none.” Sessions can also resume for up to 30 minutes, which helps with calls that pause and restart.
Grok Voice 3.0 Pricing and Access
xAI priced Grok Voice 3.0 the same as its predecessor. Speech-to-speech calls cost $0.08 per minute. xAI’s own model documentation confirms that rate. That holds steady at $4.80 per hour of audio. It matters most for any team running a contact center at volume.
Text-to-speech on its own costs $15 per million characters. Speech-to-text runs $0.10 per hour in batch mode or $0.20 per hour streaming. Access requires an xAI API key and a paid account. The model ships with closed weights, so self-hosting is not possible.
Grok Voice 3.0 Benchmark Performance
xAI leans on Artificial Analysis for its benchmark claims. Grok Voice 3.0 scores 86.9 on that firm’s Speech-to-Speech Index. No other model on the leaderboard scores higher right now. That is up from 81.3 for Grok Voice 2.0, a meaningful jump in one generation.
The gains are not universal, though. Time to first audio slowed from 0.70 seconds to 0.81 seconds between versions. It still beats rivals like GPT-Live-1, though, which answers in 1.34 seconds. On tau-Voice, a benchmark for agent task success, Grok Voice 3.0 scores 71.6%, trailing GPT-Live-1’s 74.5%. Big Bench Audio puts it at 98.5%, behind StepAudio 3 Realtime’s 99.7%.
These numbers come from xAI’s own citations of Artificial Analysis, not an independent audit. Developers evaluating voice models should treat vendor-reported benchmarks as a starting point. Testing with real call scripts still matters more than any leaderboard rank.
Migrating From Grok Voice 2.0
The upgrade did not wait for anyone to opt in. xAI pointed the grok-voice-latest alias at Grok Voice 3.0 on October 14. Any session built on that alias switched versions with no code change on the developer’s end. Most teams got no deploy to review and no warning email beforehand.
xAI has deprecated Grok Voice 2.0 but has not published a retirement date. Teams that want stability should pin the dated model ID, grok-voice-3.0, instead of the alias. A quick repository search for grok-voice-latest finds every place that needs review.
Before switching for good, run identical test calls against both versions using the same session settings. Compare completion rates, escalation counts, and minutes per call, since pricing stays flat at equal usage. Running deliberate rollouts like this fits the discipline DevX has covered in guides to integrating AI into existing development workflows. Unannounced model swaps often cause more damage there than the upgrade itself.
Common Mistakes When Adopting Grok Voice 3.0
Teams often assume unchanged pricing means unchanged behavior. It does not. Latency rose slightly. Some agentic benchmarks also dropped compared with GPT-Live-1. A straight swap can change call outcomes even at the same cost.
Skipping the SIP and compliance checklist causes a second common mistake. Grok Voice 3.0 supports SOC 2 Type II and HIPAA coverage with a signed business associate agreement (BAA). That coverage only applies if a team configures its account for it. xAI says it neither keeps call recordings nor uses them to train future models. Regulated teams should still confirm that in a signed agreement, not a blog post.
Relying on the -latest alias in production remains the most avoidable error of all. A CI check that blocks any -latest model ID from shipping prevents this. Routine updates then cannot change call behavior mid-quarter without review.
Key Takeaways
- Grok Voice 3.0 launched October 14, 2026, as xAI’s new default speech-to-speech model.
- Pricing stayed flat at $0.08 per minute for speech-to-speech calls.
- It tops Artificial Analysis’s Speech-to-Speech Index at 86.9, but latency rose to 0.81 seconds.
- The
grok-voice-latestalias auto-upgraded existing sessions without a deploy. - xAI deprecated Grok Voice 2.0, so pin
grok-voice-3.0directly for stability.
Frequently Asked Questions About Grok Voice 3.0
What Is Grok Voice 3.0 Used For?
Grok Voice 3.0 powers real-time voice agents for phone support, contact centers, and in-app assistants. It handles both sides of a spoken conversation without a separate text-to-speech step.
How Much Does Grok Voice 3.0 Cost?
Speech-to-speech calls cost $0.08 per minute, the same rate as Grok Voice 2.0. Text-to-speech and speech-to-text run on separate pricing, billed per character or per hour.
Does Grok Voice 3.0 Replace Grok Voice 2.0 Automatically?
Yes, for any integration using the grok-voice-latest alias. xAI switched that alias to point at Grok Voice 3.0 on October 14, 2026. Unpinned sessions upgraded automatically, with no code change required.
Is Grok Voice 3.0 Open Source?
No. Grok Voice 3.0 ships with closed weights, and xAI only offers it through its hosted Voice Agent API. xAI does not support self-hosting.
What Languages Does Grok Voice 3.0 Support?
The model supports 20 languages with automatic detection during a call. It can switch languages mid-conversation without a manual configuration change.
Is Grok Voice 3.0 HIPAA Compliant?
xAI says the model is HIPAA eligible when a customer signs a business associate agreement. It also lists SOC 2 Type II and GDPR compliance, though regulated teams should confirm coverage contractually.
Final Thoughts
Grok Voice 3.0 gives developers a higher-ranking, similarly priced option for voice agents. The upgrade also reached existing integrations automatically through a shared alias. That default switch is the real story for engineering teams. Test call quality before trusting vendor benchmarks, and pin a dated model ID rather than a moving target. Teams building or maintaining voice agents should audit for grok-voice-latest references this week. Run comparison calls against both model versions, then decide deliberately instead of inheriting a model change by default.
A seasoned technology executive with a proven record of developing and executing innovative strategies to scale high-growth SaaS platforms and enterprise solutions. As a hands-on CTO and systems architect, he combines technical excellence with visionary leadership to drive organizational success.
