Anthropic Builds a Way to Catch AI Companies Swapping Models

Morning, {{first name | folks}}! AI is moving into new territory this week. Anthropic is finding ways to catch AI companies secretly switching models, Google is making medical AI more accurate, and OpenAI is taking its models into physical product design. Meanwhile, Brave is facing a copyright battle over its AI content deals, and Netflix is bringing AI into the search for something to watch.

Anthropic has released FlashProve, an open-source system that could make it harder for AI providers to secretly replace expensive models with cheaper ones. It lets companies prove that a response came from the exact model a customer requested, without revealing the model's private weights.

The system was tested on Meta's 400-billion-parameter Llama 4 Maverick, with proofs verified in 17 seconds on a regular CPU. But generating them takes 70–95 times more computing power than running the model normally. Anthropic has released the code and research, though the technology is still a long way from everyday use.

Google Research and DeepMind have improved MedGemma, their open collection of medical AI models, with a new training approach that cut patient record retrieval errors by 50%. In one experiment, fine-tuning raised accuracy from 86.3% to 93.6%, helping the model find relevant information in electronic health records more reliably.

The research also introduces MedSigLIP, a model designed to interpret medical images across fields such as radiology, dermatology and pathology. Both models are available for developers to adapt, but the results come with a caveat: the record retrieval tests used simulated patient data, not real hospital records.

OpenAI is pushing AI into mechanical and electrical engineering, using its models to design components, create circuit boards and run simulations. In one experiment, two employees used GPT-6 and Astra to design a working device that connects to Codex in under a week. A manufacturer then produced a batch of prototypes in less than a month.

OpenAI is also working with Chip Ganassi Racing to turn 3D scans into editable engineering designs and with Mazda to explore vehicle aerodynamics and crash safety. The collaborations show how AI is moving beyond software and into the design of physical products, although engineering teams still need to review the results for real-world use.

The New York Times and other publishers have sued Brave, accusing the browser company of distributing their copyrighted articles to AI companies including Anthropic, Perplexity, Mistral, xAI and Cohere, without permission. The lawsuit also claims Brave has generated more than $100 million from the practice, citing third-party revenue estimates.

The complaint also takes aim at Brave's ability to bypass publisher paywalls and replace their advertisements with its own. The publishers want the company to disable the feature, end its AI content deals and pay damages. Brave now faces a challenge on two fronts: how it makes subscription content available to users and how it supplies that content to AI companies.

Netflix is testing AI Discover, a new tab on its TV app that helps viewers find something to watch based on their mood. Users can pick themes like “Something Light,” “Scare Me” or “Family Movie Night” and get recommendations that match what they're looking for.

Viewers can also tell Netflix what they liked or which titles they don't want to see again. The feature remembers that feedback and adjusts future recommendations. Netflix is starting with a limited US rollout for English-speaking members on supported TVs, with plans to expand in the coming weeks.