Live news

ChatGPT dots will pick models per task as OpenAI works to add GPT-6 Astra and GPT-6.1 Terra, Tibo says. He pitches Terra as faster and a way to widen access; no launch date given.

Apparent cyberattacks disrupted supermarket fridges and freezers in Sweden, Belgium, Luxembourg and Estonia, forcing some stores to discard food, AP reports.

Replit’s Amjad Masad tells subsidised AI subscribers to expect a “rug pull”, quoting Musk’s Grok Bot cost warning. Asked about Replit, he says: “We never subsidized.”

OpenAI plans to rename data centres “AI compute campuses” in a post-midterm policy push, Axios reports.

Ukraine’s Air Alert app is gradually recovering after a cyberattack; alerts and push notifications are back, but map-target display remains paused, its operator says. False nuclear and chemical warnings were sent on Nov 1.

Russia’s Emergencies Ministry says a cyberattack let intruders send false information through its mobile warning app. It says repairs and service restoration are under way; the notice names no attacker.

Grok API text from seven models will carry invisible watermarks by default from Nov 7, including structured output and tool arguments. Team admins can opt out per model where permitted; detection access requires an application.

From Nov 2, grok-imagine-image-quality redirects to grok-imagine-image-2.0 at low quality. The slug keeps working; docs say the replacement costs $0.01 less per image at every resolution.

Musk claims rival supercompute projects are missing their goals while SpaceX is already meeting its own, quoting praise for Colossus 2. He names no missed targets in the post.

Michigan will pilot a real-time, data-driven prison classification tool at selected men’s facilities from Dec 1, MDOC says. Developed with Recidiviz, it informs security levels that control movement and access to activities.

Claude Code on Opus 5.5 now steers away from recreating someone else’s specific game, Anthropic’s Addy Osmani says. Original games and your own designs remain supported; he invites reports of mistaken refusals.

DeepSeek’s MIT-licensed Super-agent prototype handles authorised GitHub and Feishu work across isolated task sessions. It’s research code, not an official product; parts of the accompanying design remain unimplemented.