Google Open-Sources DiffusionGemma, Boosts Generation Speed 4x

Google DeepMind released DiffusionGemma, an experimental open model using text diffusion to generate entire text blocks at once, achieving up to 4x faster output on dedicated GPUs with native vLLM support.

Google DeepMind has open-sourced DiffusionGemma, a 26B-parameter experimental language model built atop the Gemma 4 backbone that abandons traditional autoregressive generation. Instead of predicting text token by token, it employs text diffusion — generating entire 256-token blocks simultaneously and iteratively refining them through a denoising process. The result is up to 4x faster output on dedicated GPUs, with independent benchmarks showing over 1,200 output tokens per second on a single H100 at batch size 1. The model is released under an Apache license and is natively supported in vLLM — the first diffusion language model to achieve first-class inference framework integration. Sundar Pichai called it a racehorse, noting it can also self-correct and format complex Markdown in real time, dramatically reshaping expectations around inference throughput and the economics of serving large language models at scale.

xAI Launches Grok Voice API, Pricing Far Below Competitors

Grok Voice API, built on the same tech stack, supports multilingual voice agents and real-time search with human-like speed, tone, and temperature, while priced at a fraction of competitors.

xAI has launched the Grok Voice API, enabling developers to build voice agents that speak, think, and act with state-of-the-art performance. The API supports multilingual capabilities, tool calling, and real-time data search, with pricing set at a fraction of competing services. The model delivers natural timing, warmth, and expressiveness that closely mimics human conversation, positioning xAI to compete directly with ElevenLabs and other voice AI providers. The announcement follows the broader trend of frontier labs exposing their core capabilities through developer APIs rather than keeping them locked within consumer products. Early adopters in financial services are already integrating the API for real-time market sentiment analysis.

Anthropic CEO Calls for Narrowing AI-Policy Gap, Unveils Three Initiatives

CEO Dario Amodei noted AI development outpaces policy-making, and launched three new initiatives to bridge the gap between technology and governance.

In a newly published essay, Dario Amodei argues that AI is advancing at a pace policymaking institutions were never built for, and the growing chasm between technological progress and regulatory response has become the central challenge of the field. The essay lays out a concrete framework for closing this gap, drawing on Anthropic's experience developing frontier models and engaging with governments worldwide. Anthropic simultaneously announced three new initiatives spanning research transparency, policy engagement, and institutional partnerships. The move signals growing recognition across the AI industry that technical capability alone cannot address the governance questions posed by increasingly powerful models, and that labs must proactively participate in shaping the regulatory environment rather than waiting for it to be imposed from above.

Claude Fable 5 Hands-On: Slow, Pricey but Formidably Capable

Simon Willison tested Fable 5 for 5.5 hours: it has a big model smell, with 1M context and 128K max output, but costs $10 per million input tokens.

Simon Willison published his initial impressions of Claude Fable 5 after approximately five and a half hours of testing. The model carries a significant price tag — $10 per million input tokens and $50 per million output tokens — and is notably slower than comparable alternatives. Yet its raw capability is formidable: it possesses a 1-million-token context window, a 128,000-token maximum output, and knowledge current through January 2026. Willison notes the model crunches through essentially everything thrown at it, exhibiting what he calls a big model smell. Compared to Opus 4.8, Fable 5 shows richer knowledge of open-source projects and their internals. Anthropic describes the model as matching Mythos 5 performance but with stricter safety guardrails applied, which has become the subject of heated industry debate.