Skip to content

Grok Voice 3.0 at the Same Price: What Gets Done After the Conversation

AI

TL;DR

SpaceXSI releases Grok Voice 3.0 at the unchanged voice rate of $0.08 per minute. Support simulations improve and Starlink reports deployment at scale, but product value depends on completed service actions and repeat calls.

Grok Voice 3.0 at the Same Price: What Gets Done After the Conversation

SpaceXSI has released Grok Voice 3.0, giving developers a new model at the existing voice rate. I would use this improvement for after-sales services that can be completed during a call. If the agent answers better but account adjustments and replacement equipment still enter another queue, the support process has not become shorter.Release

The official announcement is dated 2026-10-14, without a time or timezone. This article uses an October 15, 08:53 Taipei cutoff and explicitly draws on the previous day’s release within the 48-hour window. SiliconANGLE updated its report at 18:27 EDT on October 14, equivalent to 06:27 on October 15 in Taipei. Reporting across midnight does not mean the model first launched today.Date Independent reporting

The model is available in the Voice Agent API at the unchanged audio rate of $0.08 per minute. Text inputs incur separate charges, so this is not the full cost of a support service. Applications using grok-voice-latest have automatically moved to the new version, with the API unchanged and 2.0 deprecated. Existing developers do not have to change interfaces, but should recognize that their model alias now points to a different version.Pricing Release notes

A support benchmark checks the database outcome

Artificial Analysis uses simulated customers in τ-Voice: agents follow policies and call tools, then the final database state is checked. Grok Voice 3.0 scores 71.6%, versus 56.5% for its predecessor. GPT-Live-1 (Astra, medium) remains ahead at 74.5%. Results average three trials where available, but that GPT-Live-1 configuration has only one, limiting confidence in the small gap.Method and results

Separately, Big Bench Audio measures average time to first audio at 0.81 seconds for the new model, versus 0.70 seconds previously. Better support-task scores therefore do not come with faster initial responses in that test; these measurements also come from different benchmarks.Latency scope For calls requiring an account adjustment, I would be more willing to accept this tradeoff. A product designed for quick lookups cannot simply inherit the same choice.

SpaceXSI says Starlink uses Grok Voice for 20,000 multilingual customer calls daily, resolving about 70% without human intervention. An agent built on 3.0 verifies customers, troubleshoots hardware, issues credits and arranges replacement equipment. These are vendor-reported operating figures. They neither establish that the new version created the entire seventy-percent result nor provide separate outcomes for Taiwan or individual languages.Deployment

The case gives voice agents a concrete use: after a customer describes faulty equipment, the system can proceed with the service request. A correct answer alone cannot establish that equipment has shipped. Suppose an agent promises a replacement but no order is created. The conversation may sound successful, yet the customer may still call again. The product should connect its commitments to completed backend actions, rather than treating a disconnected call as a closed case.

One minute rate, different service costs

A same-price upgrade reduces the price barrier to trying a stronger model. It does not guarantee that each completed service becomes cheaper. Asking several more questions to establish the request may raise audio charges; avoiding a transfer or repeat call may reduce total effort. This is an inference from the billing model and service process. The available evidence does not support a net-benefit estimate for either effect.Billing basis

I see more promise in voice products that complete defined after-sales actions than in products that merely turn existing support text into speech. Starlink’s example supports the existence of a deployment in that direction. However, the announcement gives neither the rate of repeat calls about the same issue nor the time between a request and its actual completion. Without improvement in those outcomes, a higher simulation score cannot yet be translated into less waiting for customers.Case limitations

The cover reuses an existing SpaceX brand image from this site, not the new voice interface or a Starlink call center. The event image could not be downloaded because the download environment could not resolve its hostname.

Sources:

More on this topic

Random article