Quick Answer
Grok Voice 3.0 is a speech-to-speech model now available through SpaceXAI’s Voice Agent API at $0.08 per minute of input audio. The release preserves the prior listed rate, upgrades the grok-voice-latest alias automatically, and carries strong vendor-reported benchmarks. Developers should verify usage volume, test interruption handling, and explicitly select the new model if they need migration control.
Key Takeaways
- Grok Voice 3.0 became available through the Voice Agent API on October 14, 2026.
- Input audio costs $0.08 per minute, or $4.80 per hour, at the listed rate.
- The grok-voice-latest alias now routes sessions to Grok Voice 3.0 automatically.
- SpaceXAI reports under-one-second latency and a 97.0% Full Duplex Bench score.
- Grok Voice 2.0 is deprecated, so developers should plan a deliberate migration.
What is Grok Voice 3.0?
Grok Voice 3.0 is a speech-to-speech model for building voice agents through SpaceXAI’s Voice Agent API. SpaceXAI announced the release on October 14, 2026, and positioned the model for conversations that require spoken input, spoken responses, interruptions, pauses, and rapid turn-taking. SpaceXAI’s launch announcement also says the model is available immediately.
Speech-to-speech systems process spoken audio and return spoken audio without requiring a developer to separately assemble every part of a conventional voice pipeline. The practical goal is a more natural exchange, particularly when a caller changes direction, speaks over an answer, or pauses before finishing a request. A model that responds quickly is useful only if it also handles those conversational moments without ending the turn too early.
Grok Voice 3.0 is an API product rather than a consumer app feature. That distinction matters because businesses and developers, rather than ordinary Grok users, are the immediate audience for the release. Readers looking for a separate communication feature should note that Grok’s XChat group-chat plans cover a different product area and do not describe the Voice Agent API.
How much does Grok Voice 3.0 cost?
Grok Voice 3.0 costs $0.08 per minute of input audio, or $4.80 per hour, according to SpaceXAI’s October 14 announcement. SpaceXAI lists the same input-audio rate for Grok Voice 3.0 that it listed for Grok Voice 2.0, so the new model does not introduce a higher published input price at launch.
The rate is relevant because voice-agent spending rises with total audio volume. A team running a small pilot can estimate costs from recorded or expected caller minutes, while a call center needs to account for the full amount of audio sent to the model across all conversations. The stated price covers input audio in the supplied launch details, so developers should confirm the complete current billing terms before setting a production budget.
| Usage measure | Listed Grok Voice 3.0 rate | What the figure helps estimate |
|---|---|---|
| 1 minute of input audio | $0.08 | Cost of a short spoken interaction |
| 60 minutes of input audio | $4.80 | Cost of 1 hour of incoming audio |
| Grok Voice 2.0 comparison | Same listed input rate | Whether a migration changes the published input price |
Grok Voice 3.0’s price should not be treated as a complete comparison with every competing voice platform because the supplied information does not establish equivalent billing units, output-audio charges, or feature bundles for other services. The most sensible approach is to compare a representative workload across vendors before selecting a production provider.
Will the grok-voice-latest alias upgrade automatically?
The grok-voice-latest API alias now points automatically to grok-voice-3.0, which means existing sessions using that alias receive the new model without an API-name change. SpaceXAI’s official release notes describe the alias change and identify Grok Voice 2.0 as deprecated.
Automatic routing reduces migration work for teams that intentionally use a latest-version alias. At the same time, automatic upgrades mean behavior can change without a code deployment from the customer. Voice systems are sensitive to timing, interruption handling, prompts, and escalation logic, so a team should test the updated experience against its own call flows even when the integration continues to function.
Grok Voice 2.0 users can migrate by selecting grok-voice-3.0 directly or by using grok-voice-latest. Selecting the versioned model gives a team more explicit control over which named model it calls. Using the latest alias prioritizes access to the current version, which is useful for teams prepared to validate updates as they arrive.
How do Grok Voice 3.0’s benchmarks compare?
Grok Voice 3.0 posted an 86.9 score on the Speech-to-Speech Index, compared with 83.2 for GPT-Live-1 (Astra, Medium) and 82.6 for Gemini 3.8 Live with extended thinking, according to Artificial Analysis figures cited by SpaceXAI. The Artificial Analysis speech-to-speech leaderboard and methodology provide the independent benchmark context for those scores.
Grok Voice 3.0’s index result suggests a strong showing in the benchmark’s evaluated speech-to-speech tasks, but a benchmark score is not a guarantee that every deployment will sound equally natural or solve every customer request. Accent coverage, conversation design, source data, call routing, and human escalation policies can all affect a voice agent’s practical performance.
| Benchmark or measure | Reported result | What it evaluates |
|---|---|---|
| Speech-to-Speech Index | 86.9 | Speech-to-speech performance in the Artificial Analysis comparison |
| GPT-Live-1 (Astra, Medium) | 83.2 | Comparison score cited by SpaceXAI |
| Gemini 3.8 Live with extended thinking | 82.6 | Comparison score cited by SpaceXAI |
| Full Duplex Bench | 97.0% | Pauses, interruptions, and turn-taking |
SpaceXAI also reports a 71.6% result on the τ-Voice customer-support benchmark and says the result is comparable to GPT-Live-1 at 40% lower latency. Those are vendor-reported performance claims, so teams should treat them as useful evaluation signals rather than a substitute for testing their own scripts and callers.
Why does under-one-second latency matter for voice agents?
Grok Voice 3.0 responds with under-one-second latency, according to SpaceXAI, and that speed matters because long pauses make a voice system feel less responsive during a normal conversation. A caller who receives a quick spoken reply is more likely to continue the interaction naturally, especially when asking a short follow-up question or correcting a previous instruction.
SpaceXAI says Grok Voice 3.0 scored 97.0% on Full Duplex Bench, a measure of how naturally a model handles pauses, interruptions, and turn-taking. Full-duplex conversation means both sides can manage an exchange more like people do, rather than waiting rigidly for one turn to end before the next begins. The benefit is most relevant for support, scheduling, and information calls where users often interrupt an answer.
Low latency does not remove the need for careful agent design. A rapid response can still be unhelpful if the system misunderstands the caller or continues speaking when a human wants to correct it. Developers building real-time AI experiences, including products such as camera-guided voice assistance, should measure both response speed and whether the system gives callers a clear way to interrupt, repeat, or reach a person.
Where is Grok Voice 3.0 already being used?
Grok Voice handles millions of Tesla in-car conversations per day, and Starlink call centers handle 20,000 calls daily with the system, according to SpaceXAI. The company also says roughly 70% of Starlink calls handled by Grok Voice are resolved without a human agent. Those figures indicate that SpaceXAI is presenting the model as infrastructure already used at substantial scale.
The reported resolution figure matters because a voice agent is valuable only when it completes a useful task or collects enough information for a human to continue efficiently. A system that answers routine questions can reduce the number of interactions requiring an agent, while a system that mishandles unusual requests can create extra work later. The supplied information does not define how SpaceXAI measures a resolved call, so the 70% figure should be read as the company’s own operational claim.
Grok Voice 3.0’s existing deployment examples are not a guarantee that the model fits every organization. A car conversation, a satellite-service support call, and a medical, financial, or emergency interaction have different error consequences. Organizations handling sensitive or high-stakes requests should retain human escalation and independently review the system before relying on automated spoken responses.
How should developers migrate to Grok Voice 3.0?
Developers should test Grok Voice 3.0 before treating the automatic alias upgrade as a completed migration. SpaceXAI’s release notes allow teams to select grok-voice-3.0 directly or use grok-voice-latest, while Grok Voice 2.0 is deprecated. The practical decision is whether a team values an explicit version choice or automatic access to the current named model.
- Review whether the existing integration calls grok-voice-latest, grok-voice-2.0, or another configured model name.
- Select grok-voice-3.0 explicitly if version control is more important than automatic upgrades.
- Run representative conversations that include pauses, interruptions, corrections, and transfers to a human agent.
- Compare response timing and task completion against the existing voice workflow before expanding usage.
- Monitor input-audio minutes so the $0.08-per-minute rate aligns with the team’s expected budget.
Developers should stop the rollout and contact SpaceXAI support if the updated model causes failures in production call routing, authentication, payment handling, or required human escalation. A voice agent can be tested internally, but an organization should not make a system the only route for customers who need urgent or account-specific assistance.
Who should consider Grok Voice 3.0?
Grok Voice 3.0 is most relevant for developers and organizations that need a real-time spoken interface and want to evaluate an API model with a published $0.08-per-minute input-audio rate. The automatic grok-voice-latest upgrade also makes the release relevant to existing users who may receive the new model without changing their API calls.
Teams that prioritize interruption handling may find the Full Duplex Bench result worth investigating, while teams with large call volumes may focus first on the published input rate and reported latency. The central limitation is that SpaceXAI’s operational claims and several performance figures remain company-reported, while benchmark outcomes cannot replace a test with a specific customer population and workflow.
Consumers looking for a general-purpose voice creation tool have different needs from API developers. A comparison of free AI voice generator limits can help distinguish consumer voice tools from a speech-to-speech agent platform built for application integration.
FAQ
What is Grok Voice 3.0?
Grok Voice 3.0 is a speech-to-speech model available through SpaceXAI’s Voice Agent API. SpaceXAI released the model on October 14, 2026, for voice-agent conversations that require spoken input and output.
How much does Grok Voice 3.0 cost?
Grok Voice 3.0 costs $0.08 per minute of input audio, or $4.80 per hour. SpaceXAI lists the same input-audio rate for Grok Voice 3.0 that it listed for Grok Voice 2.0.
Does grok-voice-latest now use Grok Voice 3.0?
The grok-voice-latest alias now routes automatically to Grok Voice 3.0. Developers using that alias should test their conversations because model behavior can change even when API calls do not.
Is Grok Voice 2.0 still available?
Grok Voice 2.0 is deprecated, according to SpaceXAI’s release notes. Developers can migrate by selecting grok-voice-3.0 or by using the grok-voice-latest alias.
How fast is Grok Voice 3.0?
Grok Voice 3.0 responds with under-one-second latency, according to SpaceXAI. The company also reports a 97.0% Full Duplex Bench score for handling pauses, interruptions, and turn-taking.