AI gets cheaper, more grounded—and closer to the real world

Muse Ember is being positioned as Meta’s cheapest AI model. If the low price holds up in real use, builders may find it easier to experiment with agents that make many smaller model calls.

TriNetX has introduced a tool that pulls confidence-scored facts from clinical notes using natural language processing. For health teams, the promise is easier analysis of unstructured records—but confidence scores will need to be interpreted carefully in clinical workflows.

OpenEvidence says its new Access Intelligence suite is available within its clinical AI platform at no cost. Free access could lower the barrier for clinicians to try these tools, while making clear evaluation and responsible use especially important.

Google DeepMind has unveiled a new generation of Gemini Robotics, described as its most capable robotics model so far. For developers, progress here matters when AI needs to interpret instructions and act in physical environments, not just produce text or images.

DeepMind also announced a next-generation gene prediction model. The source gives few details, but the launch is a reminder that AI’s scientific applications extend well beyond chat—and that researchers will want to understand how these predictions are validated.

Adobe’s latest creative tools include AI agents and closer integration with Google and Microsoft products. For creative teams, the practical question is how much these additions smooth handoffs and routine tasks while keeping artists in control.

Google’s Gemini Decisions API is designed to evaluate questions rather than generate responses, with reporting on support for text, images, audio, video, and PDFs and a price of $0.10 per million tokens. That could suit workflows where sorting or assessing inputs matters more than composing an answer; developers will want to test accuracy and costs on their own workloads.

WIPO launched a new edition of its Green Technology Book focused on solutions for restoring climate-impacted ecosystems. For practitioners, a curated resource can help surface technologies to investigate, though local fit and implementation remain the hard parts.

Nvidia has introduced GR00T 2, an open model for helping robots learn complex physical tasks, alongside simulation tools and benchmarks. The combination could give robotics teams more ways to develop and assess systems before putting them into real-world settings.

Data.co says its private AI platform offers one API key to reach 91 datasets from 25 providers, spanning areas such as public filings, economic indicators, and case law. For builders, consolidating access may simplify prototypes—but the usefulness will depend on data quality, coverage, and terms.