Anthropic’s Claude Is Starting to Make Its Mark on Drug Discovery

Morning, {{first name | folks}}! Anthropic is testing Claude on drug discovery, Google has a new approach to AI memory, and OpenAI is bringing ChatGPT into Oracle’s business software. Meanwhile, Google has unveiled its latest world-generation model, and Microsoft’s new image models are now generally available. Let’s get into it!

Anthropic says Claude generated around 35,000 antibody designs in a month for a potential food allergy treatment. Lab tests identified 23 promising designs, including three that appeared to dislodge IgE, a molecule involved in allergic reactions, more effectively than a standard medication.

The work is still preclinical, but other projects have shown similarly striking results: Claude helped improve a cancer therapy design and cut drug-manufacturing document review time by 89%. Anthropic is now working with six life sciences organizations to speed up different stages of drug development.

Google Research has introduced Sleep, a framework that lets language models absorb new information and retain it beyond a single conversation. In tests, the approach handled contexts of up to 10 million tokens and improved performance on tasks including factual recall, translation and mathematical reasoning.

Sleep gives models an offline phase to turn temporary information into longer-term knowledge, then practice using AI-generated examples. The goal is to help models improve after deployment without losing skills they already have, though the results are still research findings rather than a ready-to-use product.

OpenAI and Oracle are bringing business data and workflows into ChatGPT, so employees can investigate costs, understand purchasing decisions and prepare month-end reports without jumping between separate applications. Oracle is also adding OpenAI-powered tools for building apps and finding security vulnerabilities.

Oracle says its recruiting team has cut market research preparation from two to four days to just 15–20 minutes using ChatGPT Work. The partnership shows where enterprise AI is heading: beyond answering questions, toward handling work inside the systems companies already rely on.

Google DeepMind has unveiled Genie 4, a model that creates interactive worlds from text prompts. Instead of watching a generated clip, users can move through a scene, change the weather, add characters and hear matching sounds as the environment responds in real time.

The technology could let game studios test environments before building them, help filmmakers experiment with scenes and give robotics teams virtual spaces to train in. Genie 4 is still in research preview, but it points to a different use for generative AI: creating worlds people can interact with, not just images they can look at.

Microsoft has made MAI Image 2.6 and its faster Flash version generally available in Microsoft Foundry. The models generate and edit images, with finer control over text, layouts, visual styles, and changes to existing pictures.

Teams can use reference images and web information to guide outputs, making it easier to maintain consistent designs and refine specific details. Flash targets faster, lower-cost generation, while both models are available for developers to integrate into their own applications.