xAI has released its latest voice model, Grok Voice Think Fast 2.0, claiming it to be their most capable speech-to-speech model. The model aims to improve response speed and accuracy in voice interactions, potentially powering applications like virtual assistants.
SynthePulse Insight · AI deep reading
Version 1 · 1 source
xAI releases Grok Voice Think Fast 2.0, claiming significant improvements in intelligence, transcription accuracy, and conversational ability, with several benchmark figures. However, these figures largely come from third-party Artificial Analysis, and some improvements (e.g., 10x transcription boost in noise) lack independent verification. Users should evaluate based on their own scenarios.
xAI released Grok Voice Think Fast 2.0 on July 29, 2026, officially calling it the 'most powerful voice-to-voice model,' emphasizing improvements in intelligence, transcription accuracy, and conversational ability. The model is the successor to Grok Voice Think Fast 1.0 and is scheduled to automatically become the default version of grok-voice-latest on August 5, 2026, with no manual upgrade required.
Officially, the new model shows 'meaningful improvements' in voice reasoning, conversational ability, and tool-use reliability, but specific details are not provided. These statements are official claims and lack independent verification.
According to Artificial Analysis benchmarks, Grok Voice Think Fast 2.0 achieves an overall voice quality index of 82.9%, higher than the previous generation's 75.7% and GPT-Realtime-2.1's 79.1%, but lower than Gemini 3.1 Flash's 95.7% (note: this figure is for the conversational dynamics sub-item; the overall index is not listed). On speech reasoning (Big Bench Audio), the new model scores 97.2%, nearly flat with the previous generation's 97.1%; on agentic performance (τ-voice Bench), it scores 56.5%, significantly higher than the previous generation's 52.1% and GPT-Realtime-2.1's 45.7%.
In terms of transcription accuracy, xAI's official evaluation shows that across thousands of phrases in 24 languages, the new model improves 1.5–2.0x over Deepgram Nova 3 and ElevenLabs Scribe v2, and 1.4x over the previous generation. In noisy environments, the gap widens to about 10x. However, these data come from xAI's own evaluation, with no specific WER values provided and no third-party replication.
Regarding inference efficiency, the official claim is that the new model uses only 0.4x the relative reasoning tokens (P50) of the previous generation during parallel inference, with tool calls typically executed before the end of the first sentence. This data also comes from the official source and lacks independent verification.
xAI states that through reinforcement learning, the new model tends to use shorter sentences, ask only one question at a time, and avoid redundancy in conversations, thereby improving user experience. The official description says the model can guide complex workflows in the background and think ahead multiple steps, but this is a product feature description without objective measurement.
For commercial validation, xAI conducted A/B testing in Starlink customer service scenarios, claiming 'significant improvements' in sales conversion and support resolution rates, but without providing specific data or statistical significance. The test was conducted internally by xAI, and results may be influenced by various factors, so they should be interpreted with caution.
Grok Voice Think Fast 2.0 is priced at $0.08 per minute of audio, with the official emphasizing that pricing is 'predictable and transparent.' Existing users will automatically upgrade on August 5 without any action; to retain the old version, they must pin grok-voice-think-fast-1.0 before that date.
The official claims that the new model improves performance in almost all use cases without requiring prompt changes, but this statement is based on internal testing and lacks external verification.
This report is based on xAI's official blog, where benchmark data comes from third-party Artificial Analysis, but the official does not provide complete testing methodology; transcription accuracy and inference efficiency data are xAI internal evaluations lacking independent replication. Commercial validation (Starlink A/B test) does not disclose specific data. All performance claims should be treated as vendor claims, and actual effectiveness should be verified by users.
Grok Voice Think Fast 2.0 shows progress on several benchmarks, especially in inference efficiency and transcription under noise, but key data often come from official or a single third-party source, lacking complete methodology. Pricing is transparent and migration is simple, but enterprise users should conduct small-scale testing in their own scenarios before making decisions.
Primary report
Primary source