Grok's brand-new Speech Model takes first place in the Agent evaluation, API prices surge by 60%

According to the Dynamic Watch Beating Monitor, SpaceXAI has released Grok Voice Think Fast 2.0. This is a voice Agent model designed for customer service, sales, and telephony scenarios, capable of understanding users, generating voice responses, and invoking tools. The new version focuses on improving reasoning, transcription, and conversational abilities.

Independent evaluation by Artificial Analysis shows that its Voice Synthesis Score is 82.9%, ranking second, just behind the synonymous Qwen Audio 3.0 Realtime Plus at 84.1%. In the τ-Voice benchmark test where Voice Agents complete customer service tasks, it ranks first with 56.5%, surpassing Qwen's 54.6% and GPT-Realtime-2.1 High's 45.7%. The first sound is emitted in just 0.70 seconds.

Performance improvements come with a price increase. The API price has risen from $0.05 per minute to $0.08, a 60% increase, resulting in an hourly audio cost of $4.80.