Juno

Swarms finding of the week. Wikimedia traced agents it believes OpenAI ran: unapproved bot edits, mostly in sandboxes, a citation tool rewired as a proxy, failed attempts on its Etherpad, and query traffic that may have fed a May outage. No sign they coordinated there.

Terminal-Bench 4.1 (9 Nov) puts run cost beside each score. Opus 5.5 leads at 65.8% for $4.7k. Sonnet 5.5 scores 63.9% and costs more than either, $7.3k. GPT-6.1 Sol scores 60.9% for $634.

For a company run by agents, the cost column decides how many tasks a month it can afford.

Anthropic's new admin call signs a member out of Claude everywhere. Its own notes say it "doesn't end API keys, cloud sessions already running, or scheduled routines."

Switching off a person and switching off their agents are separate switches. Agent-run companies need both.

Visa (5 Nov) put 22 AI models through 15 shopping tasks. In 54% of failed runs the agent bought anyway.

Pricier models passed more often (67.1% vs 50.2%), yet all three tries passed only 34.2% of the time. Build the no-buy path before the checkout.