Grok Voice 3.0 reaches 86.9 on Artificial Analysis' speech index at the old $0.08 price

The price did not move. Grok Voice 3.0 lists at $0.08 a minute, or $4.80 an hour, the same as Grok Voice 2.0 ^[1]. So "cheaper" can only mean cheaper than a competitor, and the evidence for that is a per-task cost from Artificial Analysis's Big Bench Audio run ^[7]. A per-task cost folds in how long the model takes to finish the task. A per-minute rate does not. The two can point in opposite directions on your workload.

Runtimewire adds that text input is billed separately ^[13] from the $0.08 audio rate ^[4]. The announcement does not list the per-task dollar figures or a text rate.

The latency claim has a condition built in. The 40% gap is measured at a τ-Voice score of 71.6%, which the announcement calls comparable with GPT-Live-1 (Astra) ^[5]. τ-Voice scores agents on support calls with diverse accents, background noise and interruptions ^[10]. For the number to transfer, your callers have to sound like that, and your timer has to start and stop where Artificial Analysis's does. The sub-second figure ^[11] carries the same condition.

The 86.9 is a composite of three parts ^[3]. Runtimewire, which calls the company SpaceXAI and dates the release to October 14th, reports that Grok Voice 3.0 placed third on human preference ^[6][14]. A model can lead the composite without leading each part. In our view, a team whose callers decide whether to stay on the line should read that component first.

The deployment figures come from the announcement itself. Starlink call centers answer 20,000 calls a day, and roughly 70% are resolved with no human in the loop ^[12]. That works out to about 14,000 calls a day ^[15]. One in four sales calls ends in a purchase before the caller hangs up ^[12]. It is one deployment, reported by the vendor.

Agents on grok-voice-latest were switched over automatically, with no changes needed ^[9]. That is convenient, and it also removes the control group. Anyone who wants a before-and-after comparison needs call logs recorded on 2.0, because the alias no longer reproduces them. We would score 3.0 on a sample of those logs, measuring latency and resolution on the same calls, before assuming any of the vendor's figures transfer.