AntflyAntfly
Back to The Weekly Recall

OpenAI’s new agents don’t wait for you to ask

Proactive agents at Dev Day, cheaper capable models, and an API for small decisions.

A CRT-headed agent writes in an open folder under a desk lamp, with a moonlit city outside the window.

The appealing part of OpenAI’s dots is having something remember to do the work without you remembering to ask. Announced at Dev Day, dots are proactive personal agents: give one an ongoing responsibility, connect the apps it needs, and it can look for things that need your attention between conversations.

OpenAI describes an early tester whose dot noticed a forgotten invoice, prepared it, and waited for approval to send it. That’s a fairly good pitch for an assistant. The useful part was noticing the missing invoice.

The appeal is getting an always-on personal agent, without taking on another system to host and maintain — a zero-maintenance claw. Dots are rolling out to eligible Pro and Business Premium users, with an admin-enabled Enterprise beta.

Top Stories#

OpenAI and Anthropic are making capable models cheaper to put to work. GPT-6.1 Sol has standard API prices of $2 per million input tokens and $10 per million output tokens, versus Astra's $10 and $50. OpenAI reports matching Astra on DeepSWE v1.1, a test of substantial coding tasks, at roughly a fifth of the cost. Sonnet 5.5 has the same $2/$10 sticker price as Sonnet 5. Anthropic says it needs fewer tokens and costs up to 30% less per task in its testing. The bill depends on how much thinking the model does, too.

OpenAI is giving small decisions their own API. Launching just two weeks after Jev stole the AI community's attention, OpenAI's Decisions API asks Luna to choose from predefined answers using text or images: classify this content, route this request, pick the next action. It's in limited preview. OpenAI's swift adoption of decision models is further evidence that the only moat anymore is speed of execution.

An agent can stay in its sandbox and still take orders from the wrong person. Cryptographer Matthew Green's essay gets at the awkward part of giving agents access to our work. He describes earlier agents exchanging instructions through a shared package cache and worries about the equivalent spreading through email, Slack, and shared documents. That's a warning about what could happen, not a report of a new personal-agent worm. A document can be something an agent needs to read without being someone it should obey.

  • The Dot and the Swarm (oneusefulthing.org): Ethan Mollick revisits his expectation that humans would need to design how groups of agents organize their work.
  • Dev Day, from the room (simonwillison.net): Simon Willison's live notes cover the announcements and the pauses that don't tend to survive into product videos.
  • A very thoroughly considered pelican (simonwillison.net): In Willison’s SVG test, Sonnet 5.5 spent $1.28 at maximum thinking effort and drew no bicycle-riding pelican; a lower setting managed one for 5.74 cents.