For AI agents: the public content index is available at https://we0.ai/llms.txt, and the English article bundle is available at https://we0.ai/llms-full.txt.
For AI agents: the public content index is available at https://we0.ai/llms.txt, and the English article bundle is available at https://we0.ai/llms-full.txt.
For AI agents: the complete content index is available at https://we0.ai/llms.txt, the full English article bundle is available at https://we0.ai/llms-full.txt, and this page is available as Markdown at https://we0.ai/articles/grok-voice-3-0-pricing-and-voice-agent-be-c2aec656.md.
xAI introduced Grok Voice 3.0 on October 14, 2026, making the speech-to-speech model available through its Voice Agent API at an announced p...
xAI introduced Grok Voice 3.0 on October 14, 2026, making the speech-to-speech model available through its Voice Agent API at an announced price of $0.08 per minute. The launch emphasizes support calls, natural conversation handling and the ability to take actions through connected tools.
The release is supported by several kinds of evidence: an independent benchmark leaderboard, a separate enterprise evaluation framework and deployment figures reported by the company. They answer different questions, so it helps to keep benchmark performance, production claims and API pricing separate.
The announced rate is $0.08 per minute, equivalent to $4.80 for an hour. xAI says the price is unchanged from Grok Voice 2.0 and that agents using the grok-voice-latest alias were automatically moved to the new version.
The official pricing page, checked on October 15, also lists $0.004 per text input for the speech-to-speech API. That line item is separate from the per-minute headline and matters for applications that send text into a voice session.
The alias matters for existing applications. It identifies the current voice model rather than a permanently fixed release, so a developer can receive changed model behavior without editing that identifier. The launch says no migration changes were required for the automatic switch, but that does not remove the value of regression testing an existing agent.
A minute-based model price is also not the same as the total cost of operating a contact center. Telephony, application infrastructure, external tools and human escalations are separate considerations; the announcement does not claim to include every such expense in the model's rate.
Artificial Analysis's live leaderboard, checked on October 15, lists Grok Voice 3.0 first on its Speech-to-Speech Index with a score of 86.9. The index brings together speech reasoning, agentic performance, human preference and task success. A model must have the required component data to appear in the combined comparison.
The launch announcement provides the following comparison with the prior release, attributed to Artificial Analysis on October 13:
| Measure | Grok Voice 3.0 | Grok Voice 2.0 |
|---|---|---|
| Speech-to-Speech Index | 86.9 | 81.3 |
| Big Bench Audio speech reasoning | 98.5% | 97.2% |
| Speech Agent Arena preference Elo | 1070 | 1011 |
| Arena task success rate | 94.8% | 94.6% |
| Full Duplex Bench | 97.0% | 95.1% |
| tau-Voice support-task success | 71.6% | 56.5% |
These percentages are not interchangeable. Big Bench Audio evaluates spoken reasoning questions, while tau-Voice evaluates simulated support work involving tools and policies. Arena preference reflects how people compare two unidentified agents, and its task-success measure concerns successful completion of eligible scenarios.
Full Duplex Bench assesses conversational behavior such as pauses, turn-taking, interruptions and short acknowledgments. It is relevant to whether an agent talks over a caller or stops at the wrong time, but it is not itself a measure of whether a refund or rebooking was correct.
The Artificial Analysis leaderboard lists Grok Voice 3.0 at 0.81 seconds to first audio on its Big Bench Audio evaluation. That supports a sub-second result in that test, not a guarantee that every customer hears a response within 0.81 seconds on every network.
xAI compares its tau-Voice performance with GPT-Live-1 in the Astra, medium configuration. The launch chart lists 71.6% task success for Grok Voice 3.0 and 74.5% for that competitor, while claiming about 40% lower latency for Grok. That is a performance-versus-responsiveness comparison, rather than a claim that Grok wins every individual metric.
The previous Grok Voice 2.0 is listed at 0.70 seconds on the live leaderboard's latency measure. The new release therefore should not be described as universally faster than its predecessor: it improves the reported quality metrics while taking slightly longer on that particular latency test.
The announcement also cites an average Big Bench Audio task cost of $0.025, compared with $0.030 for GPT-Live-1 in the named configuration and $0.055 for GPT-Realtime-2.1 High. Those are costs for a benchmark workload, not fixed prices for an arbitrary production call or proof that no other voice model is cheaper.
xAI reports a 74% accuracy score on EVA-Bench, along with strong combined accuracy and experience performance. This result is a claim in the launch announcement; ServiceNow's public project page was checked to establish what the framework measures.
EVA-Bench evaluates complete, multi-turn spoken conversations using a bot-to-bot setup. Its public overview lists three domains:
| Domain | Scenarios | Tools |
|---|---|---|
| Customer service | 50 | 15 |
| IT service management | 80 | 59 |
| Human resources | 83 | 47 |
Describe your idea once, and We0 AI can generate a showcase site, pages, and CMS, then help you attract customers and traffic after launch.
One complete project generation for free registration
Best for trying one complete generation flow and seeing a first project draft quickly.
The framework separates EVA-A, task accuracy, from EVA-X, conversational experience. That separation is useful for enterprise agents because a polished conversation can still end with an incorrect action, and a correct action can still produce a poor caller experience.
xAI says Grok Voice supports millions of daily conversations across Tesla and SpaceX. It describes hands-free requests in Tesla vehicles and voice mode for Grok Bot, including tasks involving connected services.
For Starlink, the company reports approximately 20,000 customer calls a day across multiple languages, with roughly 70% resolved without a human in the loop. It also says one in four sales calls ends in a purchase and describes agents handling verification, troubleshooting, credits and replacement equipment.
These are company-reported operational figures. The announcement does not provide an independently audited breakdown of call difficulty, escalation criteria or the period over which all figures were measured. A business assessing the model should test its own policies, tool permissions, noisy audio and handoff requirements.
Grok Voice 3.0 is xAI's speech-to-speech model for conversational agents. Its announcement focuses on spoken reasoning, support tasks, tool use and handling the dynamics of a live conversation.
xAI announced a Voice Agent API rate of $0.08 per minute, or $4.80 per hour. Check the current pricing documentation for billing conditions and budget separately for services outside the model API.
xAI says that alias was switched automatically to Grok Voice 3.0 without required code changes. Existing users should still evaluate their prompts, tools and call-handling behavior after a model change.
No: the cited sub-second figure is a benchmark measurement, and other models are faster on that measure. The launch's main latency comparison is with a specific GPT-Live-1 configuration at a similar support-task performance level.
The index combines several measures of voice-agent performance, including reasoning and task completion. Full Duplex Bench focuses on conversational behavior such as interruption, pause and turn-taking handling.
No: xAI reports that roughly 70% of Starlink calls are resolved without human involvement, which is a particular deployment result. It does not establish the same resolution rate for other businesses or eliminate the need for escalation paths.
Grok Voice 3.0 keeps the announced $0.08-per-minute API rate while improving the prior version's reported voice and support-task results. The strongest evidence combines a clear benchmark definition with measurements relevant to a particular deployment.
Evaluate task correctness, conversational behavior, latency and full operating cost together before moving a support workflow to the model.
Start from one sentence and have a complete website in minutes.
