Codex Sites , Microsoft models , Anthropic cost backlash
Codex Sites , Microsoft models , Anthropic cost backlash Most teams approach AI adoption backwards (Sponsor)The question isn’t “which tool has the best model?” It’s “which solution will our team actually use?” This Notion guide breaks down the 5 critical jobs AI should solve at work and how to evaluate tools for adoption and integration, not just capabilities. Get the guide →Headlines & LaunchesBuilding a hill-climbing machine: Launching seven new MAI models (5 minute read)Microsoft released seven new MAI models, enabling developers to tune model weights themselves and integrate these into everyday products. The models leverage Frontier Tuning, an approach where AI adapts to specific workflows through reinforcement learning environments. Microsoft also announced a collaboration with Mayo Clinic to develop an advanced AI healthcare model, combining clinical expertise with AI capabilities, initially deploying within Mayo before wider distribution through Azure Foundry.MiniMax promises M3 weights after 1M-context model launch (2 minute read)MiniMax will release the model weights and a technical report for its M3 model within the next 10 days. The new model is currently available through MiniMax Code, token plans, and an API. It has a 1M-token context window and a guaranteed 512,000-token minimum for API use. The model is the first open-weight model to combine frontier coding, native multimodality, and a 1M-token context window. MiniMax lists standard API pricing up to 512,000 input tokens at $0.60 per million input and $2.40 per million output.Codex new Capabilities (6 minute read)OpenAI released new Codex capabilities and six role-specific plug-ins for data analytics, creative production, sales, product design, equity investing, and investment banking.Deep Dives & AnalysisThe Next Frontier of Visual AI Is Code (11 minute read)Visual AI is shifting from generating final pixel outputs to creating source code for editable artifacts, transforming workflows in design and 3D modeling by enabling continuous iteration and feedback. Code-native generation produces structured representations such as HTML/CSS or Blender scripts, facilitating precise edits and enhancements post-generation. This approach holds promise for industries requiring consistent 3D structures and interactive assets, with visual AI moving towards creating adaptive, editable digital artifacts rather than just static images.Open and closed models are on different exponentials (8 minute read)A lot of businesses want to switch to open models, but these models aren’t yet good enough in out-of-distribution tasks. However, they will eventually catch up. The open model ecosystem will be far more diverse and numerous than the closed lab oligopoly. The total market value of the open model economy will dramatically exceed the cumulative value of OpenAI and Anthropic.Memory Is Purpose (15 minute read)Sentra CEO Ashwin Gopinath argues memory is not a sidecar to intelligence but the layer that decides what reality the reasoning operates on, with knowledge being what was present and memory being the subset of the past that should survive because it changes future behavior. The boulder problem illustrates that the same artifact becomes different memory for sales, product, legal, engineering, and CEO views, so freezing ontology at ingestion traps the system inside a frame that is prematurely right.Engineering & ResearchPreventing AI Inference Theft at Scale (5 minute read)Vercel outlined how attackers resell stolen AI inference by exploiting exposed endpoints and highlighted why traditional rate limits are often insufficient. The post describes an approach that verifies every AI request using BotID analysis to reduce abuse.The Data Center Moves to Your Machine (4 minute read)Perplexity unveiled a hybrid local-cloud inference system at Computex 2026 that intelligently routes queries between on-device models for lightweight tasks and cloud-based models for complex reasoning, building on the company’s earlier Personal Computer agent.Wall Attention (GitHub Repo)An attention mechanism that improves long-context reasoning by organizing information around persistent “wall” memory tokens.MiscellaneousAnthropic faces AI spending backlash before IPO (3 minute read)Anthropic filed for an IPO amid growing corporate scrutiny over high AI costs, which threatens its revenue as companies reevaluate their AI investments. Businesses, Anthropic’s key clients, express concerns over spending, with a survey revealing 40% experiencing cost savings below 10%. Despite Anthropic’s strong performance, including surpassing OpenAI in business clients, a shift towards cheaper models or open-source alternatives poses a risk.Anthropic expands Mythos to 150 additional organizations in more than 15 countries (3 minute read)Anthropic has expanded Project Glasswing to an additional 150 partners in more than 15 countries. The new partners will need to meet security requirements before gaining access to the model. Since launch, Project Glasswing partners have discovered more than 10,000 high or critical-level security flaws. Major Project Glasswing partners include Apple, Nvidia, Microsoft, CrowdStrike, and Palo Alto Networks.GitHub’s plan for Agents (90 minute read)GitHub has been the home for software for almost two decades. Coding agents have seen massive quantities of code being shipped on the platform, which has grown 1,400% this year. This has put pressure on GitHub’s infrastructure, which was designed around human developers moving at human speed. This article features an interview with GitHub COO Kyle Daigle where he discusses how AI has changed things in the company, how he uses AI to work at GitHub, why the AI era is breaking GitHub in new ways, and more.Quick LinksState of Memory in Agent Harness (12 minute read)Mem0 surveyed memory implementations across Claude Code, Codex, Copilot, OpenClaw, Hermes, Bedrock AgentCore, Windsurf, and Devin, and found the same boundary failures everywhere: bounded local storage, mostly keyword retrieval, harness scoping, weak staleness handling, and 57-71% cross-user contamination rates.TinyFish Bigset turns text prompts into live datasets (3 minute read)TinyFish released Bigset, an open-source system that converts text prompts into structured datasets from the live web.Get the most interesting AI stories and breakthroughs delivered in a free daily email.Join 920,000 readers for one daily email