A new voice model behind an existing API alias
The Grok Voice 3.0 announcement makes the speech-to-speech model available through the Voice Agent API. The listed price is $0.08 per minute, unchanged from Grok Voice 2.0.
The publisher says agents using the grok-voice-latest alias have already switched to the new model without application changes. That is operationally relevant for teams whose production configuration follows the alias rather than an explicitly pinned version.
Separate benchmark claims from your call flow
The announcement reports an Artificial Analysis Speech-to-Speech Index score of 86.9 and a Full Duplex Bench score of 97%. It also cites enterprise voice-agent evaluation results and claims sub-second latency. These figures are reported by the publisher with named benchmark attribution; Franklin has not reproduced the tests.
The release describes deployment in Tesla vehicles and Starlink support. Its operational counts and resolution rates are company claims. They do not establish the containment rate another support team should expect with different customers, policies or connected tools.
A voice agent's useful result depends on more than whether it responds quickly. It must interpret the request, apply the correct policy and perform only the actions its workflow permits. A benchmark result provides evidence about a defined test, while a real service still needs evaluation against its own completion criteria.
Re-test the alias after the upgrade
Franklin analysis: teams using the latest-model alias should replay representative calls, including interruptions, noisy audio and a case requiring handoff. Compare both the spoken response and the actual downstream action.
Keep the model identifier, prompt and tool configuration with each result. At the same advertised per-minute rate, total cost still depends on call duration, retries and the work that requires a person afterward. The release is a reason to test those outcomes, not to assume its benchmark ranking predicts every customer-service workflow.