25. 9. 20263 min read

Cheaper models, faster gatekeepers, and Meta's agent hardware

OpenAI halved inference prices with GPT-6 Sol and Luna, a 100ms decision model started guarding agent context, and Meta put a personal agent on a pin.

Signals worth your attention

A 100ms decision model makes yes-or-no work almost free

Prajwal Tomar fed 468 of his own tweets, engagement numbers hidden, to Jev, a model that decides rather than writes, and got 2,340 binary judgments in 60 seconds for 1.3 cents, as reported on X. He then paired it with Claude Code as a guardrail on destructive commands, a fraud check, an inbox triage and an applicant scorer, as reported on X. Cheap classification changes what you bother to automate.

A 100ms decision model makes yes-or-no work almost free
Subscribers
This issue is for subscribers

Free, unlock with your email. You get the rest of this issue now and every next issue on Friday morning.

Already in?
Cheaper models, faster gatekeepers, and Meta's agent hardware — AI Native Weekly