The frontier is starting to run its own research

Four Sep. 6 signals point to a frontier that is starting to work on itself.

OpenAI says agents now produce 3.1 agent-workdays of research for each human workday inside its research organization. The same lab says chain-of-thought monitoring is harder to trust as models use tools, manipulate reasoning, and reason without words. 12

CNBC reports that Anthropic, OpenAI, Meta, and Google released model updates in one week, leaving buyers with more comparisons to run and less time to run them. 3

The policy signal is concrete too: EU tech chief Henna Virkkunen says US protections arrive through courts and state law, while the European Commission is checking frontier-model providers. 4

The common thread is a loop. Capability speeds up the work that builds capability, while evaluation, monitoring, and public rules have to keep pace.