DeepSeek open-sources a coding agent prototype and says human attention is the new bottleneck
DeepSeek describes its Super-agent, an always-on AI agent built around a router, and releases a research prototype on GitHub.
DeepSeek describes its Super-agent, an always-on AI agent built around a router, and releases a research prototype on GitHub.
SpaceXSI will watermark API text from seven Grok models by default, unless a team confirms a legal exception applies.
A UK government survey of 3,503 firms found the share using AI doubled in a year; 64% of users saw no change in costs.
GPT-6 Astra found a proof of a 2002 conjecture, and a test of 200 random open conjectures measures how often GPT models propose answers.
OpenAI will merge ChatGPT and Codex limits into one token-based allowance from December, with no exact date or per-plan amounts yet.
Cisco Talos details an infostealer that hands victim triage to an AI model, with two backup models in case the first one refuses.
September 2026 was Stack Overflow’s first month under 1,000 new questions since 2008, and 70% of its survey respondents ask an AI agent first.
GitHub measured a six-month shift, Microsoft is rebuilding GitHub’s Git core for agents, and Google says AI speeds up its own model work.
Four older Claude models will mark their text starting November 3. Most users can neither turn the mark off nor check it.
Anthropic’s $200 plan gets a new name and, once an account is verified, no 5-hour limit. The weekly cap stays where it was.