1. Google Undercuts Anthropic's Flagship Pricing by Half with Gemini 3.1 Pro Google released Gemini 3.1 Pro on February 19 at $2 per million input tokens and $12 per million output tokens. Claude Opus 4.6, Anthropic's top model, charges $5 and $25.

2. Microsoft Copilot Quietly Read Confidential Emails for Weeks An enterprise IT admin configured everything correctly. Sensitivity labels marked certain emails as confidential.

3. CEOs Spend $250 Billion on AI While Productivity Gauges Flatline Ninety percent of firms report zero productivity impact from AI over three years. CEOs collectively spent more than $250 billion on the technology in 2024 alone.

In Brief

• World Labs Raises $1B, Locks In Autodesk as Strategic Partner Fei-Fei Li's World Labs closed a $1 billion round, including $200 million from Autodesk. The two companies will integrate World Labs' spatial intelligence models with Autodesk's 3D tools, starting with entertainment workflows. TechCrunch

• Perplexity Drops Ads, Betting Users Won't Trust a Chatbot That Upsells Perplexity pulled advertising from its AI search product. The decision splits the industry: OpenAI is moving toward ads while Perplexity and others are banking on subscription revenue and user trust. The Verge

• Zhipu AI Releases GLM-5, Claims Parity with Frontier Models on Agentic Tasks Zhipu AI published GLM-5, a foundation model built around agentic reasoning and coding. The model uses a new architecture called DSA to cut training and inference costs while maintaining long-context performance. A new asynchronous reinforcement learning pipeline handles post-training. Hugging Face

• Amazon Kills Blue Jay Robotics Project After Six Months Amazon shut down Blue Jay, a robotics initiative that launched less than six months ago. The company said the underlying technology will feed into other robotics efforts, and affected employees moved to different teams. TechCrunch

• Google Ships Lyria 3 Music Generation Inside Gemini App Google added music creation to the Gemini app via Lyria 3. Users can generate 30-second tracks from text or image prompts. Google Blog

• Anthropic Bans Using Subscription Credentials for Third-Party Tools Anthropic updated its terms to explicitly prohibit routing Claude subscription authentication through third-party applications. The policy targets developers who build wrappers or proxies on top of consumer Claude accounts. Claude Code Docs

• Paper Finds Sparse Autoencoders May Not Beat Random Baselines Researchers tested whether Sparse Autoencoders — a popular method for interpreting neural network internals — actually recover meaningful features. Two complementary evaluations showed SAEs struggling to outperform random baselines on downstream tasks, despite strong results on synthetic benchmarks. Hugging Face

• Simon Willison: AI Coding Agents Made Him Finally Embrace Type Hints After 25 years of resisting strong typing in Python, Willison says coding agents changed the calculus. Type hints slowed him down when he typed everything himself; now that agents write the code, explicit types improve correctness without costing him iteration speed. Simon Willison

Read the full edition →