Plus: Anthropic hires Tino Cuéllar and Google DeepMind predicts cyclones.
VOICE MODELS
🎙️ Grok Voice Think Fast 2.0 reasons while speaking
Image source: x.ai
xAI launches Grok Voice Think Fast 2.0 on August 5, 2026. The speech-to-speech model reasons during queries without adding latency and costs $0.08 per minute of audio processed.
The details:
Users must pin grok-voice-think-fast-1.0 before August 5, 2026, to avoid automatic updates to version 2.0.
On August 5, 2026, grok-voice-latest will automatically update from version 1.0 to 2.0.
xAI evaluated the model across 24 different languages, measuring both intelligence and transcription accuracy.
Grok Voice Think Fast 2.0 is a next-generation speech-to-speech voice model with improved intelligence and transcription accuracy.
Why it matters: Reasoning while speaking can make voice conversations feel natural without adding a pause. Yet xAI reports a roughly 10× performance gap in noisy settings versus dedicated speech-to-text models, limiting where the integrated approach is useful today.
⚖️ Anthropic hires Tino Cuéllar as global affairs chief
Anthropic appointed Mariano-Florentino Cuéllar as its first Chief Global Affairs Officer. The former California Supreme Court justice and Carnegie Endowment president brings deep government experience to shape AI policy strategy during a critical regulatory inflection point.
The details:
Cuéllar served on Anthropic’s Long-Term Benefit Trust since January 2026 before stepping down to accept this new executive role at the company.
He worked across three presidential administrations in federal agencies and led scholars in 20 countries at Carnegie Endowment for International Peace.
Daniela Amodei stated Cuéllar spent his career helping public institutions respond to change with thoughtfulness, pragmatism, and commitment to common good.
Cuéllar began organizing AI classes at Stanford Law School nearly a decade ago to prepare legal professionals for emerging technology governance challenges.
Why it matters: Anthropic signals that regulatory engagement now requires senior government credibility, not just technical safety research. Competitors lacking equivalent policy leadership risk falling behind as democracies set AI terms. Watch whether other labs match this institutional investment or cede narrative control to better-staffed rivals.
First reported by anthropic.com, 54h ago · anthropic.com
FORECASTING
🌀 WeatherNext predicts cyclones 15 days out
Google DeepMind says WeatherNext achieves state-of-the-art cyclone forecasting accuracy in a Nature paper. The model predicts tracks and intensity up to 15 days ahead using functional generative networks trained on global atmospheric data.
The details:
WeatherNext Cyclones generates 1,000 possible scenarios for each storm using Functional Generative Networks according to Google DeepMind.
Three-day forecasts now equal prior models’ two-day forecasts, representing roughly a decade’s worth of meteorological progress per researchers.
The system processes nearly 20 terabytes of global atmospheric data and nearly 5,000 historical storms from the IBTrACS database.
A single 15-day forecast generates in less than a minute on a TPU, up from 50 predictions produced at a time last year.
Why it matters: WeatherNext provides more than 24 hours lead time advantage over current systems. This speed helps forecasters address $1.4 trillion in economic losses accumulated globally over the past 50 years as cyclone intensity modeling improves beyond traditional methods.
