Grok Voice 3.0 launches with higher scores and unchanged pricing

Grok Voice 3.0 launched on October 14 with public API access and the same headline price as its predecessor. The speech model is aimed at interactive conversations and customer service agents. For developers already using the service, the change may have arrived before they opened the announcement.

The release notes say the latest model alias now routes sessions to Voice 3.0 automatically. Version 2.0 is deprecated, and developers who selected that version explicitly are directed to change their model setting.

The upgrade changes the model behind an existing integration

The migration guide says the endpoint, session parameters and events remain unchanged. Developers can select the version through the model query parameter on the existing realtime WebSocket connection. A versioned identifier selects a specific release, while the latest alias follows the service’s current choice.

The October 14 notice does not give a final shutdown date for Voice 2.0 or say that requests explicitly selecting it are already redirected. Its deprecation should not be described as a completed retirement. The automatic move is confirmed for users of the latest alias.

That distinction matters because Grok’s separate transcription service recently retired its older model identifier and redirected those requests. The new voice conversation notice describes a different transition.

An unchanged integration can still produce different answers, timing and tool choices. Teams should record which model a deployment actually uses, then rerun representative conversations before treating the new release as equivalent to the previous one.

Higher benchmark scores come with a latency tradeoff

The Artificial Analysis leaderboard lists Voice 3.0 at 86.9 on its combined speech index, compared with 81.3 for Voice 2.0. The index combines reasoning, agent performance, human preference and task completion. A model needs all four components to receive an index score.

Its customer service task score rises from 56.5% to 71.6%, and conversational dynamics from 95.1% to 97.0%. Those results cover different tests. The customer service benchmark uses simulated callers, tools and policy documents, while conversational dynamics covers behaviors such as interruptions and turn taking.

The same table reports first audio at 0.81 seconds for Voice 3.0 and 0.70 seconds for Voice 2.0. The new model stays below a second in that test, but is not faster than its predecessor. A stronger overall score does not make it the fastest model or the leader on every component.

These are benchmark results, not a promise about a live support queue. ByteForward has not independently run the model. A useful local comparison should keep the prompts and available tools consistent and measure completed tasks, mistaken actions and response delays separately.

Pricing needs more than the headline minute rate

The current pricing page lists both versions at $0.08 per minute, or $4.80 per hour. It also lists $0.004 per text input. All figures are in US dollars. The release therefore holds the published voice rate steady rather than introducing a price cut.

The same page lists separate charges for provider tools, including web and X searches. Teams should check the rates for the tools their agent actually uses rather than treating the voice minute rate as a complete operating budget.

For a deployment decision, cost per successfully resolved issue is more useful than the voice rate alone. Track repeat calls and human escalations alongside billed usage. A shorter call is not a saving if the customer must call back to fix the result.

Existing controls still shape the experience

The current guide documents session instructions, selectable voices, tool access and automatic or manual turn handling. It also exposes reasoning settings, vocabulary hints and pronunciation substitutions. These are documented integration controls, not all newly introduced features of Voice 3.0. Keep their settings visible when comparing versions so a configuration change is not mistaken for a model improvement.

The API reference provides a server endpoint for creating expiring client secrets for realtime use. Its example sets a 300 second expiry. This is an existing integration mechanism, not an October 14 launch feature.

For browser and mobile clients, the guide recommends short lived tokens. Long lived API keys belong on the server, rather than inside the client application.

Deployment claims are useful context, not customer guarantees

The launch announcement says Grok Voice 3.0 serves conversations in Tesla vehicles and powers Grok Bot’s voice mode. It also describes Starlink support operations and says roughly 70% of its calls are resolved without a human. Those are company reported deployment figures, not independently audited outcomes or guarantees for another business.

Before expanding a voice agent’s responsibilities, test accents, noisy audio, interruptions and requests outside its authority. Check when it hands a conversation to a person and whether it completes the intended action. The practical value of this release depends on those outcomes as well as its improved benchmark scores.

Illustrative recording equipment photographed by Anna Pou in 2021, used under the Pexels License. The photograph does not show Grok or a launch demonstration.