Latest / Minds, Bodies, and Terawatts / OpenAI Launches GPT-5.6 and 'ChatGPT Work,' an Autonomous Office Agent, and Lifts Government Access Limits

Transcript

OpenAI Launches GPT-5.6 and 'ChatGPT Work,' an Autonomous Office Agent, and Lifts Government Access Limits · Minds, Bodies, and Terawatts

• 0:00
  [Host] This is Minds, Bodies, and Terawatts, where your next coworker ships in a software update. CNBC reported on July ninth that OpenAI rolled out GPT-5.6, and quietly shipped something bigger alongside it: a tool called ChatGPT Work.
• 0:14
  [Guest] A family called Sol, Terra, and Luna, but the models aren't the story. OpenAI just stopped selling you a chatbot and started selling you a coworker.
• 0:23
  [Host] A coworker how? What does ChatGPT Work do that the old ChatGPT didn't?
• 0:28
  [Guest] The old one waited for your next question. This one takes a goal, opens your apps, and runs the whole job itself, pulling the report, drafting the email, updating the spreadsheet, then handing it back finished.
• 0:42
  [Host] And the sell to a company watching every AI dollar?
• 0:46
  [Guest] Sam Altman told CNBC the flagship, Sol, is fifty-four percent more token-efficient on agentic coding, and as good or better than rival models. More work per dollar is exactly the pitch for handing it a whole workflow.
• 1:00
  [Host] But here's the friction. The reporting says these models shipped at limited capacity, at the U.S. government's request. That's not an agent for everyone.
• 1:09
  [Guest] Right. OpenAI worked with Commerce Secretary Howard Lutnick before launch, over cyber-safety and keeping frontier models at home. It opens to the general public, but on a leash the government is still holding.
• 1:21
  [Host] So the pitch to a company is, stop hiring for the busywork.
• 1:26
  [Guest] The pitch is your existing staff each run a small team of these agents. OpenAI shipped it as a desktop app, so the thing lives on your machine, right next to the actual work.
• 1:39
  [Host] Here's my worry. If the agent runs the whole workflow, what's left for the person?
• 1:45
  [Guest] Direction. The job moves from doing the task to specifying it, and catching the agent when it's confidently wrong.
• 1:52
  [Host] That sounds like a promotion for everyone. Is it?
• 1:55
  [Guest] Only if you're good at it. Gartner found only seventeen percent of companies have actually deployed an agent, and it expects forty percent of agent projects to be scrapped by 2027, killed by cost and weak oversight.
• 2:09
  [Host] So the tooling has outrun the skill to run it.
• 2:13
  [Guest] Badly. The discipline even has a name, governance-as-configuration. You stop checking every output and start writing the rules the agent follows, so someone answers for it when a billing agent codes the profitable answer instead of the correct one.
• 2:28
  [Host] And if nobody's watching that chain?
• 2:30
  [Guest] You get the slow-motion failure. First you check everything, then you check the summaries, then the summaries are written by another agent, and no human can read the whole chain anymore. The human isn't the bottleneck. The human is the brakes.
• 2:46
  [Host] So the day OpenAI starts selling a coworker, the last human job in the building becomes deciding what that coworker is allowed to touch. Patrick Deglon's name for that division in Unscarcity is AI as referee, humans as conscience.
• 3:00
  [Guest] And he warns the real trap is the company that keeps no one around to answer for the machine, so you hold onto a minimum layer of accountable people over the agents. For more on where an autonomous agent has to stop, head to unscarcity dot ai. The book is on Amazon, the audiobook is on Spotify. This has been Minds, Bodies, and Terawatts.