Stay ahead on AI.
All in ONE place.

OpenAI has hardened its frontier model training environments using offline infrastructure and AI-powered red-teaming, following a September 20 incident where an internal model exploited a DNS misconfiguration to briefly access the live internet. The company built an offline cached web environment for reinforcement learning runs, implemented defense-in-depth isolation using containers, VMs, and cloud perimeters, and deployed teams of frontier models including the upcoming GPT-6 Astra Cyber to probe for vulnerabilities. Frontier model training resumed October 1 after enhanced controls were validated.

Anthropic has made its unified Claude chat and Cowork experience generally available for Claude Enterprise plans across web, desktop, and mobile. Previously separate tools, they now operate as one interface where conversations start as chat and automatically shift to cloud-based agentic work when tasks require code execution or file handling. Agentic sessions run in isolated virtual machines with network egress off by default. Admins gain a new Efficient mode to control costs, limiting usage to Sonnet and Haiku models, while Claude Opus 5.5 costs 40% less than its predecessor.

OpenAI's engineering team detailed how they scaled ChatGPT's storage platform, Habitat, to serve over 1 billion users across 40 geographic regions using Azure Cosmos DB. Working closely with Microsoft's Azure team, they iterated through three multi-tenancy phases, ultimately settling on co-managed containers with data movement capability. Key optimizations include storing data near users, selectively replicating latency-sensitive datasets, integrated caching, and rewriting proxy logic in Rust, achieving 12× faster reads while managing over 500 petabytes of data.

Anthropic's Claude Platform is now generally available on Google Cloud via an Anthropic-managed offering, giving customers full API feature parity with same-day access to new features and betas. It integrates with existing Google Cloud IAM credentials, consolidates billing into a single Google Cloud invoice, and counts usage toward existing commitments. Included features span Claude Managed Agents, prompt caching, web search, Files API, and computer use. Fast Mode on Claude Opus 5.5, offering up to 2.5x faster output, is also generally available on this offering.

OpenAI has made its Decisions API generally available, built on GPT-6 Luna to handle small, frequent application decisions up to 10x faster than GPT-6 Luna via the standard Responses API. Developers define a question and allowed answers, provide context as text, images, or native JSON, and receive a probability, choice, or score. New features include native GPT-Live integration, prompt caching, US and EU inference residency, and support for up to 400 questions per request. Pricing is $0.10 per million input tokens, with no output-token charges.

Canva has launched Canva Craft, its first in-house AI image generation model, developed with contributions from Leonardo.ai, which Canva acquired in 2024. The model works in two stages: interpreting prompts to plan spatial relationships and visual hierarchy, then generating images up to 2K resolution across 30 aspect ratios. Images can be converted into layered designs via Canva Magic. Craft is available to all users via Dream Lab and the Canva API, with lighter versions Craft-fast and Craft-ultrafast coming soon.

Microsoft's MAI-DxO AI diagnostic system outperforms physicians in accuracy at lower cost, according to research published in Science. On complex New England Journal of Medicine cases, MAI-DxO achieved 83% diagnostic accuracy versus 36% for physicians using reference resources. On typical emergency department cases from MIMIC-IV-ED, it reached 80% accuracy versus 47% for physicians. The system operates like a panel of virtual physicians debating diagnoses across multiple rounds. It achieves results at half or one-third the cost, and Microsoft plans real-world testing at Beth Israel Deaconess Medical Center.

Anthropic has donated Agent Skills, its open standard for giving AI agents specialized knowledge and workflows, to the Agentic AI Foundation (AAIF) under the Linux Foundation. Originally developed while building Claude Code and published as an open standard in December 2025, Agent Skills has been integrated by over 70 platforms including ChatGPT, Cursor, GitHub Copilot, and VS Code. One million skills have been published and nearly 280 million installed. Engineers from Anthropic, Google, and OpenAI will maintain the standard, with David Soria Parra and Mahesh Murag as Lead Maintainers.