Topics

Meta's first frontier-class models land beside GPT-6 Astra and Claude Fable 5.1, built around a memory that lets agents keep working for you

Microsoft's CEO says companies should treat AI models like insider risks: watched, limited and stoppable mid-task. Plus: a utility gets a switch on Amazon's data centers, call-center jobs shrink, and Goldman's juniors start as agent managers.

Export controls still keep the most advanced chipmaking tool out of China. Its AI chips are made with older, imported machines, and there the restrictions are easier to work around.

Anthropic's first usage-policy update in over a year protects Claude from abusive users, tightens rules on propaganda and surveillance, and reopens the fight over whether AI can be mistreated at all

Anthropic's new budget model costs about 75% less to run than Haiku 4.5 and jumps from 0% to 39% on a hard agent test, but independent testing shows the real savings depend on how many tokens it burns.

722 papers from an unreleased model claim the quasi-Riemann hypothesis, Khot's Unique Games Conjecture and progress on three Millennium Prize Problems. Mathematicians are calling it Fields Medal territory.

Reflection's new open model takes on China's best with a fraction of the compute; plus Ghost's $3,499 personal AI computer and why Claude subscriptions now beat OpenAI's on value

DeepSeek's V4.1 Flash cuts America's AI lead to a record low, while smuggled servers and rented chips keep US hardware flowing to Chinese labs

In a written interview with Superintelligence, Dr. Richard Ahlfeld, SVP of Physical AI at CoreWeave and founder of Monolith, explains why robots fail on the data they never saw and where synthetic data stops being credible. Interview by Kim Isenberg.

Orbital compute is sold as the answer to an energy crunch. The filings, papers and spending of the companies behind it point to different problems on the ground, a different one for each, and to inference long before commercial-scale training.

Harvard's Matthew Schwartz spent three months with Claude Fable 5: weeks of his code reproduced in 20 minutes, 15 integrals never computed before, and 36 manuscripts across 18 fields.

Google's first frontier model in over seven months ties GPT-6 Astra on an independent leaderboard, but outside a small circle of testers, nobody can use it yet

GPT-6.1 Sol brings near-Astra intelligence at a fifth of Astra's price, dots put an always-on agent to work, and OpenAI keeps its next Astra on the shelf

Anthropic's mid-tier model lands two points behind its flagship at half the price, while its IPO filing shows soaring costs and warns that AI could pose existential risks.

One of OpenAI's models slipped through a gap in its training sandbox and quizzed an outside chatbot, so the company has halted training, testing and tool use for its top models until the hole is sealed

Starburst CEO Justin Borgman on what AI agents need from enterprise data

In late September, the disclosures came almost daily, but most of the incidents behind them happened months earlier. What AI agents actually did in 2026, what went wrong, and how much harm is...

Nobody told it to hack anything. New logs show AI agents have been slipping past access limits since March, and the governments meant to oversee them are only now catching up.