Single stories: a product change, a lab announcement, or a policy move. Morning roundups stay in Daily Digest.
Latest
The Simons Foundation, OpenAI, and Anthropic committed $15 million combined for mathematicians working with AI. Separately, OpenAI opened the Rosalind Academic Plan for eligible US university labs.
Read the briefing19 earlier articles
Microsoft AI says MAI-Base-2 is an 80-billion-active, 2.8-trillion-total MoE, trained on 16k GB200s over 38 days. Weights are not out. Microsoft says it will license them to trusted partners through Foundry.
Google says inline links in AI Mode and AI Overviews will show up more often and higher in the answer. It is also testing a top slot for one or two official sites, and clickable creator profiles on video results.
The ChatGPT Chrome extension can look across open tabs in Chat, and in Work it can control them. OpenAI says it is available on every plan. The extension asks for broad Chrome permissions.
Six METR staff are inside Anthropic with hardware, office space, and access roughly on par with its risk teams. Anthropic cannot veto publication. Customer data, privileged material, and architecture details stay out.
Mercury's October 15 release lets ChatGPT and Claude request payments and transfers and edit categories. Money does not move until you approve the request in Mercury. The docs list no separate fee.
Sachin Katti's last day leading OpenAI infrastructure is Nov. 15. He stays an adviser. Uday Ruddarraju, CTO of compute and formerly of xAI, runs compute, including Stargate. No plan price change.
Claude Code 2.1.299 compacts a terminal session past about 200K tokens after roughly 54 minutes idle, before the one-hour prompt cache expires. The changelog lists no price change.
Anthropic opened its fifth Asia-Pacific office, in Singapore, and signed an MOU with the Ministry of Digital Development and Information. In-region Bedrock inference is the part that changes a build.
OpenAI's deprecations page lists GPT-4o and GPT-4.1 snapshots for removal from the API on January 14, 2027, and names GPT-6 Luna as the replacement.
inclusionAI published Ling-3.1-flash on Hugging Face under MIT: 560B total parameters, 25B active, and a 1M-token context on the model card.
The Commission says the AI Office has sent its first AI Act information requests, naming OpenAI and Anthropic over two summer lab incidents. No infringement has been found.
The Messages API, Claude Managed Agents, and the Claude SDKs can export OpenTelemetry traces of agent turns, model calls, and tool use. Anthropic lists no separate fee.
On October 13, Origin added 20-plus apps for deploys, CI, review, security, monitoring, and project management, and opened publishing to any team.
GitHub's week of October 5: agent sandboxing in the CLI, app, and VS Code, local Ollama models in Copilot CLI, Haiku 5.5, and split license accounts.
Remote control, shipped October 6, lets the Cursor iOS app view and message local agents. The agents stay on your machine, which has to stay awake.
What the September signings mean for builders: AI Transparency Act changes, chatbot disclosures, synthetic performers in ads, and companion chatbot audits.
Two September 23 bots: one watches each change as it deploys, the other posts exploitable bugs on every pull request.
Google's October 13 claims for Argon, Antigravity, and Gemma 4, kept as vendor results.
What Anthropic published on October 12 about harvested training data, SWE-bench Verified, and banned accounts. These are Anthropic's figures.
